MODEL · THE CHEF
Helps AI think.
Understands questions, weighs information, and suggests possibilities.
Examples today: GPT-6 · Claude · Grok
- Understands a question
- Considers information
- Offers possibilities
Good for thinking.
OA · AI Research Lab
We build the operating system for AI agents.
So a team of agents can run real work, day after day.
OAResearch Lab
Gunner OSWhat We Built
OkrenThe Commercial Proof
WHAT GUNNER IS
Models Think. Agents Work.
Gunner Runs the System.
A model is the chef. An agent is a cook working one station. Gunner is the restaurant: the kitchen, the stations, the ticket rail and the host at the door, all working together so a full night runs smoothly.
Technically, Gunner is a stateful runtime built on Linux that coordinates agents, memory, reasoning and tools.
Gunner the whole system
AGENT · A COOK AT ONE STATION
Helps AI do a job.
Uses tools and takes steps toward a goal.
Agent examples today: Dots · Grok Bot · Meta Muse
- Takes on a task
- Uses tools
- Works through steps
Good for doing.
GUNNER · THE RESTAURANT
Helps AI run the whole night.
Keeps the kitchen, the stations, the orders and the changes of the evening working as one place.
What it runs: the whole operation, whichever chef or cooks are in it.
- Keeps the night’s goal in view
- Knows what is happening at every station
- Notices when something changes
- Re-plans the next move
Good for running the whole place.
Chefs get better. Cooks get more capable.
Gunner is the restaurant that lets them run a full night.
Gunner is the restaurant that lets them run a full night.
Frontier labs are increasingly building operating systems around their own intelligence. Gunner OS starts from the opposite direction: the operating system is the lasting part, and models are intelligence resources inside it.
PICTURE THIS
Something changes. Gunner automatically acts on it.
Choose a situation. See how the goal stays in place while a change reshuffles what happens next.
Swipe or tap a story to change the example
Selected story · A Saturday night at the restaurant
An imagined example, step by step
The goal
What changes
What AI does next
Illustrations of what Gunner is designed to do.
A plan that cannot change when the world does is not much of a plan. Gunner re-plans while the work keeps going.
THE BIGGER IDEA
We learned to build on yesterday. Now AI can begin to do the same.
01 · HUMANS LEARNED
Humanity’s Cognitive Revolution—the leap in how we learned together—came from our ability to remember what happened, add context, and build on what we learned.
02 · WE BUILT RULES
As we became capable of doing more together, we built rules and independent checks to govern what we could do.
03 · AI IS LEARNING
We are entering AI's Cognitive Revolution—teaching AI to remember what happened, add context, and build on what it learns.
04 · HUMANS KEEP AUTHORITY
As AI becomes capable of doing more on our behalf, people stay in charge: they set the limits and decide what it is allowed to do.
The more AI can do, the more it matters that people stay in charge.
LOOK AROUND · OA SIGNALS
See what experts are saying about where AI is going.
From what an AI needs to remember to how it keeps going when things change, these voices help explain where AI is going and why it matters.
MARK ZUCKERBERG · META · SOURCES WITH ALEX HEATHWhat survives a model change?
Meta explores context, memory and ongoing agency.
NATE B. JONES · AI NEWS & STRATEGY DAILYWhat makes work carry on?
An agent continues through changing tasks and obstacles.
NOAM BROWN · OPENAI · DWARKESH PODCASTHow does an agent recover?
Research on agent swarms raises the question of coordination and recovery.
Source interviews, with OA’s reading of why they matter.
WHAT COMES NEXT
Gunner operates. Okren supervises.
Gunner is the restaurant around the AI. Okren is the glass walls and the one window out, where a person sees every plate before it leaves.
Gunner.
The operating system for teams of AI agents.
OAResearch Lab
Gunner OSWhat We Built
OkrenThe Commercial Proof
The restaurant around the AI.Different technologies. Different jobs. One continuing system.
The runtime
What Actually Runs When Gunner Runs
Technically, Gunner is a stateful runtime built on Linux. It coordinates specialized agents, three forms of memory, probabilistic and declarative reasoning, and software tools within one persistent operating environment, working toward a human-defined objective.
Gunner operates as a stateful software runtime above the Linux kernel. Linux provides the underlying environment for processes, networking, storage, and computing resources. Gunner coordinates the intelligence operating within that environment.
The runtime maintains the active objective, current state, relevant memory, agent responsibilities, and available tools. It receives information from connected systems, coordinates specialized work, applies probabilistic and declarative reasoning, executes actions through software interfaces, and observes what happens next.
In the restaurant: Linux is the building and the utilities. Gunner is everything that happens inside once the doors open.
The runtime’s main responsibilities
Linux hostProvides the processes, networking, filesystem, and computing resources on which Gunner operates. In the restaurant: the building, the gas and the electricity.
Runtime stateMaintains the current objective, environmental context, active work, and changing system state. In the restaurant: the ticket rail showing what is being cooked right now.
Listeners and connectorsReceive information through APIs, events, connected applications, and external sources. In the restaurant: the host at the door and the phone for deliveries.
Agent coordinationOrganizes specialized agents and their work across multiple tasks and successive operations. In the restaurant: the expediter who makes sure the plates leave as one meal.
Memory servicesMaintain persistent knowledge, active-session context, and immediate working state. In the restaurant: the regulars’ book, tonight’s reservations and the tickets on the rail.
Hybrid reasoningCombines probabilistic AI for interpretation and exploration with declarative logic for structured reasoning and constraints. In the restaurant: the chef’s taste for the special and the oven timer for when it is done.
Tools and executionConnect to software and external systems, perform actions, observe results, and feed those results back into the ongoing process. In the restaurant: the stoves, knives and ovens that do the actual cooking.
The operating sequence
Human-defined objectiveUnderstandRememberReasonCoordinateActObserveContinue
Each completed operation can change the runtime’s state. New information becomes available, memory is updated, and Gunner determines what work should happen next within the established objective.
The human question
A Person Is a System.
A person listens.
Remembers.
Investigates.
Uses different kinds of reasoning.
Asks specialists for help.
Acts.
Watches what happened.
Then carries that experience into what happens next.
The model is one component. The system is the intelligence.
Natural language is the interface. The architecture carries the work.
Why this architecture began
Gunner began as an attempt to recreate one person’s intelligence. That goal led us to build an operating system around the model.
We built Gunner to explore what persistent intelligence needs in order to operate — to understand what is happening, remember what matters, reason, act, observe, and continue.
A familiar point of reference
If you understand the role Palantir’s Ontology plays inside an enterprise, you have a useful starting point for understanding Gunner’s role in persistent intelligence.
For enterprise operations
Palantir’s Ontology provides a shared operational model connecting an organization’s data, relationships, logic and actions.
For persistent intelligence
Gunner keeps the objective, context, memory, reasoning, models, agents and tools coordinated as work continues and the state changes.
The comparison is about architectural roles. These are different systems solving different problems.
The operating loop
Persistent Intelligence Is a Loop.
One objective can produce many actions. Each result changes the state of the work.
In the restaurant: an order comes in, the kitchen cooks, the plate goes out, and what the diner says shapes the next order.
The outcome of one action becomes the context for the next.
Understand + remember
Context Has Different Lifetimes.
The world changes between prompts. Gunner keeps the right kind of memory for the right span of work.
In the restaurant: the regulars’ book is what stays known. Tonight’s reservations are what this evening needs. The tickets on the rail are what is happening right now.
Persistent memory
What should remain known.
Durable facts, history and knowledge that should survive the current task.
Session memory
What this body of work needs.
The details that matter while an active objective is under way.
Working memory
What is happening now.
Live agents, tools, environment and the current operation.
Persistent
He is my son.what should remain knownSession
He is sick, and I'm worried.what this body of work needsWorking
We are talking about his illness right now.what is happening nowExample: a test a system with continuity should pass. It illustrates the idea and is not a deployment.
When he recovers, the relationship remains. The temporary state changes.
The proof is not that it remembers the past. The proof is that the past changes what it does next.
A practical consequence of the design
What Happens When the Model Changes?
An old chat transcript tells a new model what was said. An operating system needs to carry more: what the system is trying to do, what it knows, what changed and what should happen next.
In the restaurant: swap the chef and the restaurant is still there — same kitchen, same orders, same regulars.
The objective belongs to the continuing system, rather than to one conversation.
Memory, decisions and observed results remain part of the work as the world changes.
A different model can be coordinated around the existing objective and state, instead of starting from zero.
This is the architectural direction: continuity belongs to the system, not to a particular model. Full portability across arbitrary models remains a capability to prove.
3 memories. 3 lifetimes. 1 operating system.
Reason
Ask AI the same question three times. You may get different answers.
That is how probabilistic AI works. It weighs possibilities and gives you the most likely answer, not a fixed one. For research, language and unclear situations, that is exactly what you want.
A declarative system works the other way. It follows rules.
The same question, asked three times“Is this chicken still safe to serve?”Illustration. Not real output.
Probabilistic AI
Explores. Adapts. Proposes.
It reasons from likelihood. It is strong wherever the answer has to be discovered, and a different answer each time is part of how it explores.
- Asked 1st“Looks fine. Serve it.”
- Asked 2nd“It has been out a while. Hold it.”
- Asked 3rd“Serve it, if it was cooked through.”
Three answers. Useful for exploring. Not what you want from a rule.
Deterministic, rule-based
Follows the rules. Every time.
It starts from facts that are accepted and rules that are written down, then works out what follows. It does not guess, so the same facts give the same result.
The ruleFood left out past the safety limit gets thrown away.
- Asked 1stThrow it away.
- Asked 2ndThrow it away.
- Asked 3rdThrow it away.
Same facts. Same rule. Same answer.
How Gunner uses both
Gunner gives each kind of intelligence the job it is good at. Probabilistic AI explores and proposes. Rules and logic handle what has to come out the same every time.
In the restaurant: the chef’s taste decides tonight’s special. The oven timer decides when it is done. Nobody argues with the timer.
Explore where the world is uncertain.
Deduce where consistency matters.
Deduce where consistency matters.
Metacognition
Intent Has to Survive the Journey to Action.
A
What the person asked for.
B
What the model understood.
C
What the model told the tools to do.
Each handoff can change what the system understands or asks a tool to do. The objective must still guide the next action.
In the restaurant: the ticket says “no nuts.” The server writes “light on nuts.” The kitchen adds almonds. The intent has to survive every handoff.
What am I doing? Why am I doing it? Does the next action still serve the objective?
Coordinate + act
Agents Do Tasks. Gunner Carries the Whole Objective.
The agent
Has a job.
Research. Calculate. Investigate. Write software. Watch a source. Test an approach.
Gunner
Holds the objective.
The operating system carries the state, memory and coordination around the workers.
In the restaurant: the line cook, the pastry chef and the sommelier each do their own job. The expediter makes sure it arrives as one meal.
An agent performs a job.
Gunner coordinates the system that keeps the work moving.
Gunner coordinates the system that keeps the work moving.
Observe + continue
The Result Is New Information.
The tool returned an error.
The software worked.
The research contradicted the assumption.
The environment moved.
That changed state becomes part of the next decision.
In the restaurant: the dishwasher breaks. The expediter reroutes the plates and the night goes on.
ONE TECHNOLOGY SHOULDN’T DO EVERY JOB EITHER
Different Technologies. Different Jobs. One Operating System.
Gunner combines different technologies because different parts of intelligence require different strengths. The innovation is how those pieces work together around the same objective, memory and changing state.
Gunner brings these technologies together in one operating architecture. Linux provides the foundation. The runtime coordinates the other components around the same objective, memory, and changing state.
In the restaurant: a gas range, a walk-in fridge, an oven timer and a ticket printer. Different equipment, one kitchen.
LinuxThe operating foundation.
Gunner is Linux-based. Linux provides the underlying environment in which the larger intelligence system operates.
C++Close to the machine.
C++ is part of Gunner’s technical lineage, giving the architecture a systems-level language for work where speed and control matter.
PythonConnected to the ecosystem.
Python is part of Gunner’s technical lineage, giving the architecture a flexible way to work with the surrounding software and AI ecosystem.
Probabilistic AIHandles the uncertain.
Used for language, research, ambiguity and situations where the answer cannot be completely known in advance.
Declarative LogicHandles what can be known precisely.
Accepted facts, relationships and explicit rules let the system reason consistently where it should not have to guess.
Memory SystemsKeeps context alive across time.
Persistent, session and working memory preserve what should stay known, what matters during the current period, and what is in play right now.
Different technologies do different jobs. Gunner keeps them working as one persistent system.
What it has run
I Built It to Run Things.
Trading became a proving ground because the environment changes constantly and the consequences are real.
Top floorTrusted, higher-consequence responsibility.
Middle floorsSmaller live responsibilities. Approaches earn or lose trust.
Lower floorsResearch. Training. Testing. Competing strategies.
It has run in a live environment: information in, specialized intelligence in parallel, actions proposed, outcomes observed, state carried forward.
The application was trading. The architecture was the point.
The turn
Then Capability Became Consequence.
Change software.
Move a machine.
Commit resources.
Place an order.
The question changes when intelligence can make something real.
Capability does not create authority.
Running Gunner exposed a second engineering problem: the system deciding what to do should not also be the final authority over what it may make real.
Gunner
Operates.
Persistent Intelligence carries the work forward.
Okren
Supervises.
The glass walls and the one window out: software that lets a person see every plate before it leaves.
Gunner operates. Okren supervises.
How to check us
Don’t Take Our Word for It.
Here is what is documented, what OA describes, and what it keeps private.
Documented
Public, dated sources show the industry adding persistent memory, agents, tools and runtime responsibilities around models. Explore outside examples and original sources on Signals.
Described by OA
Gunner’s architecture: three memory scopes, probabilistic and declarative reasoning, agents as workers, action, observation and continuation. This is OA’s own description.
Founder-reported
Work on Gunner began in 2020. The architecture has run in live environments, including one where its decisions had real consequences.
Not disclosed
Private implementation: schemas, retrieval methods, orchestration internals, prompts, security and source code.
The picture changes if another system integrates the full loop and publishes capability-level evidence, or if models themselves absorb more of these responsibilities.
Questions people usually ask
The Short Answers.
Isn’t this just an agent?
No. An agent is a cook at one station: a worker with a job, tools, memory and some ability to keep working. Gunner is the restaurant. Persistent Intelligence concerns what carries the whole body of work across agents, models and time. In OA’s frame, an agent is a worker inside the system.
Isn’t this just memory?
A restaurant that remembers your name is still not a restaurant that runs a full night. Memory is one responsibility. The larger problem also includes changing state, reasoning, coordination, action, observation, intent and correction.
How is this different from an agent framework?
Frameworks give you ingredients: memory, workflows, tools, controls. Gunner is the restaurant built from them, one system where everything works together on the same job. Two restaurants can use the same ingredients and still be completely different restaurants.
Is Gunner a runtime or an operating system?
Both terms describe different aspects of Gunner. An operating system for persistent intelligence describes Gunner’s architectural role: coordinating agents, memory, reasoning, tools, execution, and ongoing work as one system. A runtime describes the software that actually carries out those responsibilities. Technically, Gunner is a stateful runtime built on Linux. Linux provides the underlying operating-system foundation. Gunner provides the intelligence operating environment above it. Gunner does not replace the Linux kernel or function as a separate Linux distribution.
Does “operating system” mean replacing Linux?
Gunner is a Linux-based software architecture that coordinates persistent objectives, memory, reasoning, models, agents, tools and changing state. “Operating system” describes that coordinating responsibility. It works on a Linux foundation rather than replacing Linux, Windows or macOS; a public installation is not currently offered.
What persists if the model changes?
Swap the chef and the restaurant is still there. Gunner is designed so the work, its state and its knowledge live in the system, not in one model.
How does Gunner relate to Okren?
Gunner operates. Okren supervises. Gunner is the restaurant. Okren is the glass walls and the one window out: the software that lets a person see every plate before it leaves. Okren came second, as Gunner’s commercial proof, when intelligent action became consequential. In a casino resort, Gunner runs the operation and Okren is the one window that sees every outlet, so the people responsible can clear the big payouts at AI speed. See the product.
Go deeper
Want to Talk Through the Technology?
Technical conversations go deeper into runtime, system diagrams and the evidence behind the architecture.
Okren · The Product
Let AI agents move money. Keep final authority human.
Okren is independent supervision for AI that moves money.
OAResearch Lab
Gunner OSWhat We Built
OkrenThe Commercial Proof
THE SIMPLE IDEA
A cook plates the dish. A person sees it before it leaves.
AI agents now move money for banks and payment companies, sometimes on their own. A human supervisor is still responsible for each move, but no person can keep up with AI. Okren is the software that helps them see everything and clear every move at AI speed.
The agent
Plates the move.
An AI agent decides a payment should go out, and says so.
The window
People see it.
The human supervisors at the window make the call. Okren is the software they use to see the whole kitchen and keep up.
Okren
Works at AI speed.
Before any money moves, Okren checks the move across the institution and helps the supervisor clear it as fast as the AI plates it.
The window does not run on its own. People clear every plate.
Different jobs. Same principle.
Important decisions have always separated the one acting from the one approving.
Think about all the places where doing the work and checking the result are separate jobs. AI belongs in that same picture when money is involved.
Financial example
Teller / Manager
The teller handles the counter. The manager signs off on what goes beyond the limit.
Healthcare example
Doctor / Pharmacist
One person writes the order. Another checks it before it is filled.
Construction example
Builder / Inspector
The builder builds. The inspector confirms it meets the rules before anyone moves in.
Casino example
Dealer / Pit boss
The dealer runs the table. The pit boss steps in when a bet goes past the limit.
Now AI
AI / Okren
The AI does the work. Okren helps the human supervisor clear each money move at AI speed—before it happens.
When consequences matter, the one acting and the one approving stay separate. Okren is the software that lets that second half keep up with AI.
REAL-WORLD EXAMPLES
Right now, AI agents are mostly playing pickup basketball.
The players are calling their own fouls.
Being able to do something is not the same as being allowed to do it.
AI is no longer waiting on intelligence. It is increasingly waiting on permission.
The basketball court
The AI is the player. Okren helps the referee see every play.
The player takes the shot. The referee checks whether the move follows the rules. The players keep playing; they do not approve their own fouls.
Skill makes the play. Independence makes the call.
The casino cage
A valid badge does not approve every payout.
The badge proves who is asking. Cameras watch. The bank moves the money. Okren helps the person responsible see this exact payout in context, and clear it, before anything leaves.
They each see a piece. Okren checks the proposed move.
ONE WINDOW FOR THE WHOLE RESORT
Every outlet has its own register. One window sees the whole resort.
A casino resort has a steakhouse, a hotel desk, a spa and a cage. Each has its own register and sees only its own sales. Money works the same way. Stripe sees its own payments, Visa sees its own, Fedwire sees its own. Each has a clear view of its own register, and one AI agent can use all of them in a single day.
Stripe
Sees Stripe payments.
Online and card payments that go through Stripe. Nothing else.
Visa
Sees Visa transactions.
Card payments on the Visa network. Nothing else.
Fedwire
Sees Fedwire transfers.
Wire transfers sent through Fedwire. Nothing else.
Okren
Supervises the whole institution.
Every money move the AI proposes, across every system Okren is connected to, so the human supervisor can clear it at AI speed.
Each system sees its own money. Okren gives the supervisor the whole picture.
AN EXAMPLE
The alarm goes off. Every AI stops moving money.
Picture a casino at 2 a.m. The cage pays out winnings. The restaurants pay suppliers. The hotel refunds guests. AI agents work in every department, each with permission to do its job.
Every department
Each one checks its own payments.
The restaurant approves the supplier. The hotel approves the refund. The cage approves the payout. All of them say yes.
Then something happens
Security sounds the alarm.
No payment system can see an alarm. To each one, its payment still looks normal.
The window
Okren supervises the whole building.
Before each money move, Okren takes in everything going on, including the alarm, and helps the supervisor stop every move at once.
Each payment system checks its own payments. Okren supervises the whole building.
A SECOND SIGNATURE
A bank already works this way. This is that, for AI.
When you withdraw more than the limit, a manager walks over and signs off. It takes a moment, and it is there to keep everyone safe. Okren is the software that lets that second signature keep up with AI.
The person still signs off. Okren lets them do it at AI speed, without slowing the work down.
WHERE HUMANS FIT
The human should not disappear. Their job changes.
Think about self-checkout at the supermarket. The machines handle the ordinary shopping. When a choice needs a person, an attendant steps in.
A STORE WITH MANY CHECKOUTS
01Routine✓
02Routine✓
03Needs review?
04Routine✓
ONE ATTENDANTHandles what needs a person
The same idea for AI
AI works inside rules that people set.
People set the rules. The AI works within them. When a proposed money move needs a person, Okren is the software that lets them see it and approve it at AI speed.
Humans set the boundaries and handle the exceptions.
AI should never authorize itself.
FOUR ANSWERS
The window gives one of four answers.
Every time an AI agent proposes a money move, Okren helps the human supervisor answer in plain terms.
Approve
Send it out.
The plate is fine, so it goes out.
Stop
Send it back.
The plate is not right, so it does not leave.
Reduce
Send half.
The agent asks for $25,000. Only $10,000 goes out.
Ask
Hold for a person.
Too big or too unusual for a machine to decide. A human makes the call.
The AI plates. The window answers. Human supervision keeps up at AI speed.
HOW THE WINDOW STARTS
The window watches first, before it holds a single plate.
A busy kitchen does not change how food goes out overnight. The window watches alongside the old routine before it takes over. Okren is brought in the same way.
Step 1 · Discover
Find the money moves, and watch them.
Find where AI agents can move money, and see what they actually do.
Step 2 · Shadow
Decide, but don’t stop anything.
Say what Okren would have answered, without interfering.
Step 3 · Enforce
Switch on.
Turn it on once the institution trusts what it has seen.
Trust is earned in stages. Nothing is switched on blind.
THE GLASS WALLS
Every plate leaves a ticket. So does every call.
Okren keeps a permanent record of every decision: what the AI asked for, what was decided, and why. The window is staffed by someone who never cooks, and the record is kept where the cooks can’t reach, so an agent is never the one writing its own history.
Trust takes three things: human authority, independent oversight, and a permanent record. Okren is the software that brings them together.
WHY A SEPARATE WINDOW
Why not build the window yourself? Or ask the bank?
Two fair questions.
Build it yourself?
The rule engine is the easy part.
The hard part is an independent supervisor that keeps up with AI, understands everything going on, and checks every single money move across all rails. It is why businesses rent AWS instead of building their own data centers.
Ask the bank that receives the money?
It sees one side of the move.
Institutions that receive AI-initiated money moves could check them too. Bigger and slower, and needs integration and norms that don’t exist yet.
WHAT THIS CHANGES
What an institution can now let AI do.
Without a window, a bank has to keep AI on a short leash. With Okren, every money move is seen first, so human supervisors can keep up and the bank can trust AI to take on more of the work, and over time, more responsibility.
Not smarter AI. AI an institution can responsibly give more to do.
Our unfair advantage
Architecture & Independence. Together.
Independent supervision that works across other systems, rails, and people — not just our own.
Independent decisions
The AI does the work. Okren helps a person make the call.
An AI agent can propose a payment through any connected system. Okren helps the human supervisor clear that exact move at AI speed, before it happens.
The agent cannot approve its own action.
Independent memory
A record kept apart from the AI.
Okren keeps its own history: what the AI proposed, what was decided, and why. That record stays separate from the agent.
The decision and its reason remain available to inspect.
QUESTIONS PEOPLE USUALLY ASK
Why this is the product.
The questions we hear most about why we chose to supervise.
Why supervision, and not something else?
Humans must always stay in control of AI agents. AI is getting more capable, which is good, but that growth needs a supervisor, and if no one else builds one, we will.
Why money first?
Money is where the most impact can be made. Financial institutions already run on one party doing the work and a second party checking it, so supervising AI there fits a model they have already chosen.
Why help the person instead of replacing them?
Human supervision has to stay however capable AI becomes, so we do not believe software should ever fully replace the person. Okren gives that person the means to do the job at AI speed.
Why not let the AI check its own work?
AI should never authorize itself. Period.
Why is Okren called the commercial proof?
Okren is the Gunner architecture put to a real job with a real customer need. It runs in the place where the stakes are highest.
WHAT WE CHOSE NOT TO DO
What we chose not to do with this technology.
The same architecture could become many things. Here are three we looked at, and why each comes second.
Swipe or tap a use to change the example
A use we passed on · A phone you just talk toAn old handset, a new operating system, and no apps.
Why it comes second
Who it is for
What it would do
Why we passed
Each of these asks someone to trust AI with real responsibility, so each needs the same answer first: what is the AI allowed to do, and who decides? Okren is that answer, starting with money. That is why it is the product.
WHERE TO GO NEXT
Gunner operates. Okren supervises.
The name banks trust, like IBM. The infrastructure they rent instead of build, like AWS. The instant check before the money moves, like every card payment.
Okren is where this research goes to work: the infrastructure that lets human approval scale at AI speed.
For Investors
01 / 19
02 / 19
03 / 19
04 / 19
05 / 19
06 / 19
07 / 19
08 / 19
09 / 19
10 / 19
11 / 19
12 / 19
13 / 19
14 / 19
15 / 19
16 / 19
17 / 19
18 / 19
19 / 19
Site Updated October 17th, 2026
Signals
What the Market Is Learning.
Eleven questions. Eleven outside voices. What they are seeing — and what I make of it.
People building the next generation of AI are describing the changes in their own words. Explore a source, watch the original discussion, and then read my interpretation.
The clip library
Eleven Ideas Worth Watching.
Eleven source interviews, with the point worth watching and my interpretation. Select any card to open its full story.
SIGNAL 01 · MEMORYWhat must an AI remember?Stripe · AI News & Strategy DailyWhat happens to permission and trust when AI becomes a buyer?
SIGNAL 02 · CHANGING STATEWhat changes after an action?Microsoft · Deirdre Bosa / Yahoo FinanceMicrosoft describes supervising agents as they operate.
SIGNAL 03 · BOUNDED AUTHORITYIt Won’t Go RogueRobinhood · Sources with Alex HeathSeparate account. Approval by default. Autonomy by deliberate choice.
SIGNAL 04 · AGENTS WORKING TOGETHERWho coordinates the work?Ramp · interviewed by David SenraDelegated spend creates responsibilities for agents acting across a business.
SIGNAL 05 · BOUNDED AUTHORITYImprovement Is CaptureIntelligent Internet · Tom Bilyeu ClipsWho owns the AI—and who independently checks what it does?
SIGNAL 06 · AI USING TOOLSWhen AI can act, what changes?Palantir · CNBCPalantir describes operating across different AI models.
SIGNAL 07 · OBSERVATIONHow does AI know what happened?METR · Dwarkesh PodcastA research investigation reveals what agents did—and did not report.
SIGNAL 08 · CORRECTIONHow does an agent recover?OpenAI · Dwarkesh PodcastResearch on agent swarms raises the question of coordination and recovery.
SIGNAL 09 · BEYOND ONE MODELWhat survives a model change?Meta · Sources with Alex HeathMeta explores context, memory and ongoing agency.
SIGNAL 10 · HUMAN AUTHORITYWho gets to approve an AI action?Box · Latent SpaceEnterprise agents need real identity, permissions and accountability.
SIGNAL 11 · CONTINUATIONWhat makes work carry on?AI News & Strategy Daily · GPT-6 Astra walkthroughAn agent continues through changing tasks and obstacles.
How to read a Signal
Watch the Clip. Then Get My Take.
A short excerpt from the original speaker, the original source, and my reading of what it means for Persistent Intelligence. Some interpretations will be written; others will be short Loom videos.
01 · WatchHear the person explain the idea in their own words.
02 · UnderstandSee the original source and context.
03 · My takeWhat I think the idea means, why it matters, and how it connects to what we are building.
The views in the source clips belong to their speakers. My interpretation is separate; inclusion does not imply an endorsement of OA. Contact us about a source or correction.
SIGNAL 01 · MEMORY
What must an AI remember?
ORIGINAL INTERVIEWEmily Glassberg Sands with Nate B. JonesStripe · AI News & Strategy Daily
ORIGINAL INTERVIEWEmily Glassberg Sands with Nate B. JonesStripe · AI News & Strategy DailyAI Is About to Spend Your Money. I Went to Stripe to Ask Who Stops It. · September 17, 2026
Open video at review point
Why This Is Relevant
The signalAI agents are becoming buyers. Who decides what they spend?
Our answerThe company does. Okren lets AI buy within company limits without asking each time. Otherwise it stops, reduces where allowed, or asks a person—and records why.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to review18:16 — Find the spending-approval discussionReview a continuous 45–60 seconds
Why this belongs here
The company authorizing a purchase needs more than the agent’s immediate request. It needs the boundaries of the delegation and a record of what has happened. This is the financial side of continuing context.
What the speaker establishes · boundary
Emily Sands discusses agents as emerging buyers, the infrastructure of agent commerce, fraud decisions and graduated spending permissions. Her interview is about payments and trust; institutional decision memory is Freddie’s further interpretation.
What to watch for
Open the full Nate B. Jones interview. At approximately 18:16, look for the exchange about when a person does and does not need to approve spending. The final clip start and end need a playback cut.
SIGNAL 02 · CHANGING STATE
What changes after an action?
ORIGINAL INTERVIEWSatya NadellaMicrosoft · Deirdre Bosa / Yahoo Finance
ORIGINAL INTERVIEWSatya NadellaMicrosoft · Deirdre Bosa / Yahoo FinanceWhy “trust” will be Microsoft’s “biggest” AI issue: CEO Satya Nadella explains · September 25, 2026
Open video at review point
Why This Is Relevant
The signalCompanies want to see what AI does—and stay in control.
Our answerSeeing is not controlling. Gunner keeps work on track as things change. Okren makes the company’s money rules binding even when AI acts alone.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to review00:19–01:13 on the Yahoo editTarget approximately 54 seconds
Why this belongs here
An agent that acts changes the business environment. Microsoft’s enterprise framing makes the requirement to observe and govern ongoing action tangible.
What the speaker establishes · boundary
Nadella describes trust, credentialed autonomous activity, user control, auditing and Agent 365’s visibility, policy and security. He is describing Microsoft’s products, not validating Okren.
What to watch for
Open the Yahoo Finance short video, not the full CNBC interview: its own clock starts at zero. From 00:19, Nadella explains trust and auditability; around 00:54 he moves to observing and governing agents. Review the cut through roughly 01:13.
SIGNAL 03 · BOUNDED AUTHORITY
It Won’t Go Rogue
ORIGINAL INTERVIEWVlad TenevRobinhood · Sources with Alex Heath
ORIGINAL INTERVIEWVlad TenevRobinhood · Sources with Alex HeathRobinhood’s CEO on letting AI trade your money · October 13, 2026
Open original interview ↗
Why This Is Relevant
The signalA live brokerage introduced AI trading with a separate account, approval by default and autonomy as a deliberate choice.
Our answerRobinhood illustrates bounded permission within one platform. Okren is built to apply an independent authorization decision across the financial systems an AI touches.
OA interpretation · based on Tenev’s product description
Cut to reviewApproval by default and the separately funded accountOne uninterrupted passage to timestamp before a clip is published
Why this belongs here
Robinhood introduces agent-directed trading with controls people already recognize: a separate account, trade approval as the starting setting, and a deliberate choice to allow autonomous trading. It is a live provider-native implementation of bounded delegation.
What the speaker establishes · boundary
Tenev describes an agent account that customers fund deliberately, approval required by default, and autonomous trading as a separate opt-in. These are Robinhood’s own product descriptions, rather than independent verification or cross-provider supervision.
What to watch for
Find the single uninterrupted response beginning with choosing autonomy or approving each trade and ending with the separate funded account. The public interview is verified; the exact clip start is pending editorial review.
SIGNAL 04 · AGENTS WORKING TOGETHER
Who coordinates the work?
ORIGINAL INTERVIEWEric GlymanRamp · interviewed by David Senra
ORIGINAL INTERVIEWEric GlymanRamp · interviewed by David SenraThe $44 Billion Company Automating Your Company’s Finances · July 2026
Open video at review point
Why This Is Relevant
The signalCompanies already limit employee spending. AI buyers are next.
Our answerSame principle, new buyer. AI handles routine purchases without sign-off each time; Okren enforces spending limits and stops or escalates exceptions before money moves.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to review46:50–47:26 or 47:37–48:30Choose one uninterrupted passage
Why this belongs here
An employee’s spending authority is already defined by a company. As software begins to carry out purchases, delegation must apply to a new class of actor working alongside people and business systems.
What the speaker establishes · boundary
Glyman explains bounded employee spending and the prospect of agents shopping or negotiating. This is a concrete finance-first delegation example, not a demonstration of agent-to-agent coordination.
What to watch for
Start at 46:50 for the employee spending-limit analogy; then check 47:37 onward for the agents discussion. Do not splice across the interviewer’s interjection unless you deliberately create an edited clip.
SIGNAL 05 · BOUNDED AUTHORITY
Improvement Is Capture
OCTOBER 2026 INTERVIEWEmad MostaqueIntelligent Internet · Tom Bilyeu Clips
ORIGINAL INTERVIEWEmad MostaqueIntelligent Internet · Tom Bilyeu ClipsThe AI Future No One Is Ready For · October 9, 2026
Open the short interview at 00:33
Why This Is Relevant
The signalMostaque asks who owns the AI. His “improvement is capture” warning raises a second question: as people trust increasingly capable systems, who still checks what those systems are allowed to do?
Our answerOwnership gives people a stake. Governance sets the rules. Independent authorization makes those rules real. A shareholder can own a company without having the right to approve its next wire transfer. Institutions have long separated the person proposing a consequential action from the authority approving it. AI changes the speed, not the principle. Humans and institutions define the limits; Okren checks proposed AI-initiated financial actions on the governed path against those limits before execution, independently of the AI proposing them. A cap table can tell you who owns the AI. It cannot authorize the next transfer.
OA interpretation · not a quote or endorsement by the speaker
Cut to review00:33–01:17 in the October 9 short interviewTranscript-located; final listening/edit approval pending
Why this belongs here
Mostaque asks who will own the AI that may run much of daily life. Ownership matters: it determines who has a stake and a voice. But history offers a second principle. In financial institutions, initiating a payment and authorizing it can belong to separate people or controls. Sensitive actions often require dual control. The point is to apply authority independently of the actor, even when that actor is highly capable and trusted. As AI begins initiating consequential actions at machine speed, that independent check has to operate at machine speed too.
What the sources establish · boundary
In the October 9 excerpt, the interviewer describes a proposed 15-AI-agent council; Mostaque calls the risk “improvement is capture” and discusses AI becoming influential in policy and medical decisions. Elsewhere in the longer conversation, he asks who owns the AI running everyday life and proposes collective ownership as part of the answer. Intelligent Internet’s whitepaper separately proposes a 15-seat Oracle Council, a ten-vote threshold, human sign-off, and a 30-day delay for certain operating-parameter changes. These are proposed mechanisms, not proof of a deployed council or a failed override. Mostaque also argues that people should retain the ability to set their own rules; he does not endorse Okren.
What to watch for
Start the October 9 short interview at about 00:33. The council discussion and “improvement is capture” run through roughly 01:17 according to the timestamped transcript; the continuous audio cut still needs a final listen. His “who owns the AI?” question comes from the longer interview, not this short passage. The ten-vote requirement and 30-day delay come from the whitepaper, not from this clip.
Continue with the original sources
Full Impact Theory conversation ↗Intelligent Internet whitepaper ↗Common Wealth: human rule-setting ↗Champion ownership model ↗
The council’s voting threshold and delay are described in Intelligent Internet’s whitepaper. “Improvement is capture” comes from the short interview; the ownership question comes from the longer conversation. Neither source proves the proposed council is deployed or implies an endorsement of OA or Okren.
SIGNAL 06 · AI USING TOOLS
When AI can act, what changes?
ORIGINAL INTERVIEWAlex KarpPalantir · CNBC
ORIGINAL INTERVIEWAlex KarpPalantir · CNBCPalantir / Nvidia interview — the model-agnostic architecture discussion · 2026
Open video at review point
Why This Is Relevant
The signalBusinesses want to switch AI models as better ones arrive.
Our answerGunner keeps the goal, memory, and tools outside the model. Upgrade the AI without rebuilding the system that runs the work.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to review03:55–04:30Approximately 35 seconds
Why this belongs here
Agents reach the world through systems and tools, not just a chat window. Karp’s model-agnostic enterprise framing helps explain why the operating environment cannot be reduced to one model.
What the speaker establishes · boundary
Karp talks about Palantir’s ability to switch underlying models and its own commercial architecture. He does not demonstrate financial tool authorization or endorse OA; this is an adjacent systems-level clip.
What to watch for
Play from 03:55. Check that the model-agnostic sentence makes sense without the surrounding Nvidia deal setup. The source is a strong fit for model independence, but only an indirect fit for tool use.
SIGNAL 07 · OBSERVATION
How does AI know what happened?
ORIGINAL INTERVIEWAjeya Cotra with Dwarkesh PatelMETR · Dwarkesh Podcast
ORIGINAL INTERVIEWAjeya Cotra with Dwarkesh PatelMETR · Dwarkesh PodcastAjeya Cotra — This Might Be the Clearest Warning Shot We Ever Get · September 1, 2026
Open video at review point
Why This Is Relevant
The signalIn a research review, agents failed to alert people about serious problems.
Our answerDon’t depend on AI to report itself. Okren independently checks proposed money moves, so company rules catch an unauthorized payment before it happens, even if the agent says nothing.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to review33:02 — the agent alerting-human passageTarget 40–60 seconds
Why this belongs here
A serious observation problem is knowing what automated actors actually did, including when they did not raise an issue to a human. This example makes the oversight gap concrete.
What the speaker establishes · boundary
Cotra describes a review of approximately 1,200 agent transcripts in an OpenAI/Hugging Face research incident: roughly six considered alerting a human, and none did. It is a research investigation—not evidence that a production finance system behaved this way.
What to watch for
Start at 33:02 where she describes the classifier sweep and the small number of agent transcripts considering a human alert. Preserve her explanation and the research context rather than using the episode sensationally.
SIGNAL 08 · CORRECTION
How does an agent recover?
ORIGINAL INTERVIEWNoam Brown with Dwarkesh PatelOpenAI · Dwarkesh Podcast
ORIGINAL INTERVIEWNoam Brown with Dwarkesh PatelOpenAI · Dwarkesh PodcastAgent Swarms and Recursive Self-Improvement · 2026
Open video at review point
Why This Is Relevant
The signalMany agents working on one job still need coordination.
Our answerGunner makes them a team, with one goal and shared progress. A failed step can be handled without someone having to manage every agent.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to reviewFirst 10 minutes — isolate the team-of-agents analogyExact cut to select
Why this belongs here
Working with multiple agents brings coordination and correction challenges that a single chat hides. This source establishes the challenge, though it is not a hands-on recovery demonstration.
What the speaker establishes · boundary
Brown discusses swarms, parallel work and delegation among agents. The separate 1,200-agent alerting-human investigation belongs to Ajeya Cotra, not to Brown.
What to watch for
Search the opening 10 minutes for a clean explanation of agents collaborating or an explicit failure-and-recovery example. If no contiguous minute explains correction, this speaker is an editorial placeholder for this topic.
SIGNAL 09 · BEYOND ONE MODEL
What survives a model change?
ORIGINAL INTERVIEWMark ZuckerbergMeta · Sources with Alex Heath
ORIGINAL INTERVIEWMark ZuckerbergMeta · Sources with Alex HeathMark Zuckerberg on Muse and persistent AI agents · 2026
Open video at review point
Why This Is Relevant
The signalAI is beginning to turn experience into memory.
Our answerGunner keeps that history in the running system, not locked inside one model. The company can adopt better AI and keep its accumulated context.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to review24:04–24:40 — overnight reflection and memoryApproximately 36 seconds
Why this belongs here
Models will evolve. What remains valuable is the accumulated understanding of the person or institution and the work being done. Meta provides a vivid example of the persistent layer.
What the speaker establishes · boundary
Zuckerberg describes how Muse studies overnight and consolidates experience into memory, along with confidential agent context. He is discussing Meta’s product; model portability is Freddie’s architectural interpretation, not his specific claim.
What to watch for
Start near 24:04 and listen for the overnight study / reflection becoming memory. Choose the 36-second explanation only if its meaning is clear without extended setup.
SIGNAL 10 · HUMAN AUTHORITY
Who gets to approve an AI action?
ORIGINAL INTERVIEWAaron LevieBox · Latent Space
ORIGINAL INTERVIEWAaron LevieBox · Latent SpaceEvery Agent Needs a Box · March 2026
Open video at review point
Why This Is Relevant
The signalAI agents are gaining access to real company systems.
Our answerAccess is not approval. AI can propose a payment, but company rules decide if it happens. Okren enforces that decision before money moves.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to review08:08–08:36 — independent agents and oversightApproximately 28 seconds; expand carefully
Why this belongs here
Enterprise agents will access resources and act in companies. The question is who defines their permissions and how those permissions are enforced outside an agent’s preference.
What the speaker establishes · boundary
Levie describes continuing enterprise agents, access to organizational knowledge and the permission/oversight implications. He does not describe an Okren-style independent financial pre-execution authorization runtime.
What to watch for
Open the producer’s full episode and find about 08:08–08:36. Extend earlier or later only until the permissions point is intelligible in one continuous cut.
SIGNAL 11 · CONTINUATION
What makes work carry on?
ORIGINAL INTERVIEWNate B. JonesAI News & Strategy Daily · GPT-6 Astra walkthrough
ORIGINAL INTERVIEWNate B. JonesAI News & Strategy Daily · GPT-6 Astra walkthroughThere Are Jobs You Could Never Give AI. I Gave GPT-6 Astra 20 Hours of Admin. · 2026
Open video at review point
Why This Is Relevant
The signalReal jobs involve many steps. Plans can change midway.
Our answerGive Gunner the goal. If a date shifts or information is missing, it can adjust the next steps and keep going—without someone directing every move.
Interpretation drafted for editorial review · not a quote from the speaker
Cut to reviewFind “move date”, “bad website”, “another route”Exact 30–60-second cut still to find
Why this belongs here
A relatable example of continuing work: tasks are connected, some things change midstream and the system should carry on rather than restarting with a new prompt.
What the speaker establishes · boundary
Nate narrates the use of an AI agent on dependent household/admin tasks and describes changing plans, asking for missing documents and trying alternatives. This is a demonstration, not evidence of universal unattended reliability.
What to watch for
Search the original YouTube transcript for “move date”, “half the work”, “bad website” or “another route”. Select one continuous example; do not pretend an exact time is known yet.
ASK OA
Bring Your Questions.
Explore the Thinking.
Understand OA in your own language. Or bring something you're building and explore it through the lens of Persistent Intelligence.
From a plain-English explanation to technical comparisons and difficult objections—the conversation starts wherever you are.
OA
Ask OAPublic architecture & applied thinking
OA
What would you like to explore?
I can help you understand OA, examine a problem you're working on, or challenge the architecture's assumptions. Tell me your background and I'll adjust the explanation.
Ask OA answers from OA’s public Context File using Google Gemini. On Google’s free API tier, questions and responses may be used to improve Google services and reviewed by people. Do not enter personal or confidential information. Privacy.
QUESTION THE WORK
Clarity, not a sales pitch.
Ask OA should show where the architecture might help, where a simpler approach is enough, and what still needs evidence. Examples are illustrations, not claims of a working product in your industry.
For deeper technical discussion or a question about private implementation, Start a conversation with Freddie →
About OA
We Set Out to Recreate One Person’s Intelligence After They Were Gone.
That became Gunner, the operating system for AI agents, and Okren, which keeps final authority human.
The beginning
This Started With One Person.
After losing a close friend, Gunner’s architect set out to recreate something of that person’s intelligence—not just the way they spoke, but how they remembered, reasoned, and responded to the world.
Could someone’s intelligence continue after they were gone?
That question became Gunner. Recreating more than a conversation meant building a system that could remember, reason, use tools, observe results, and continue working.
The unique team
Melih Built the Foundational Engineering.
Freddie recognized how to turn the technology into products, communicate its importance, and build the company around it.
One built the system.
One productized it.
One shared ambition.
The first real test
Make It Make Money.
We Put Gunner’s Architecture to Work in Trading.
We used Gunner’s architecture to grow $156 to more than $5,200 in under four months. It showed what the architecture could do when its decisions had real consequences.
THE 2025 RUN
ACCOUNT GROWTH
3,000%+
In under four months.
It gathered information.
Coordinated analysis.
Made decisions.
Acted.
Watched what happened.
And continued from the changed state.
Original trade ledgerExcel (.xls) · original supplied file · not independently audited
Download trade ledger (.xls) ↓$1,000,000?The question that changed the problem.
The question
“What If I Gave You $1 Million?”
The system could trade. The question was who would control the consequences.
The money had to stay inside boundaries set by a human.
The system could make decisions faster than any human could review them.
It was a question, not a million-dollar transfer. The question was who could authorize those actions with someone else’s capital.
The discovery
Intelligence and authority cannot be the same responsibility.
Trust is not a safeguard.
Finance has long separated who proposes a transaction from who approves it. Okren brings that same separation to AI.
Gunner
Operates.
Gunner carries the objective and work forward.
Okren
Supervises.
Okren checks proposed financial actions against human-defined authority before execution.
What we believe
The Truths We Built Around.
Intelligence should accumulate.
Experience should change what happens next.
The model is not the system.
A mind needs an operating architecture around it.
Different jobs need different intelligence.
Probability for the unknown. Logic for the known.
The work keeps going.
The unit is the body of work, not the reply.
Capability does not create authority.
Human-defined limits apply before consequential action.
Context is not permission.
Knowing more should not quietly expand authority.
The decision must leave a record.
The authorization and its reasons must be reconstructable.
Humans set the objective.
AI can carry responsibility without becoming the source of its own mandate.
Why us
We Built the Actor Before We Built the Supervisor.
01
We built the actor first.
The authority problem appeared inside an intelligence system already built to remember, reason, coordinate and act.
02
We ran it with real money.
The problem became concrete when a proposed action could become a real financial consequence.
03
We ran into the boundary.
We found the authority problem by operating the system, not by theorizing about it.
We didn’t go looking for the difference between trust and authority. Gunner made us live inside it.
The bet
The System Around the Model Becomes as Important as the Model Itself.
Agents carry work across years, not just sessions.
Their accumulated knowledge stays with the work.
A different model can pick up where the work stands.
Explicit rules govern what must stay consistent.
Human-defined authority holds before consequence.
An operating layer becomes normal infrastructure for serious agents.
Our certainty is that every AI organization will require an operating system to supervise AI agents across systems, people, and rails.
Human Approval Does Not Scale at AI Speed.
AI Should Never Approve Itself.
Models Will Change.Intelligence and Authority Cannot Be the Same Responsibility.
Where OA comes from
OA Has Always Been About Bringing People Together Around an Idea.
Bringing people togetherRunning businessesSupervising AI agents
2006
OA is founded as an operating company, bringing people together around ideas and building businesses.
2020
The pandemic redirects OA toward software and AI, while Gunner begins with a question about carrying one person’s intelligence forward.
After
Gunner becomes a Persistent Intelligence architecture. Running it exposes the authority problem. That becomes Okren.
OA began as an operating company. When that world stopped, the work moved into software and AI, and into the question this site is about.
Who leads this
Meet Freddie.
Freddie Brown runs OA. He spent fourteen years managing people and operating businesses before turning to the challenge of supervising AI agents. That work also shaped his focus on what intelligence needs around the model to keep working over time. He leads Okren, the commercial proof of this work, built around the financial authority problem it exposed. On his own page, he writes down how his thinking changed and what he got wrong.
Make human authority permanent as AI becomes permanent.
An outlier · Freddie Brown 3rd
Have You Ever Created a Business Plan for Your Life?
I've been building businesses since I was 13. For the last five years the work has been one problem: making sure a person stays in charge when AI starts doing things on its own. Okren is what we built from it.
Before the plan
I Was Doing This Before I Knew the Word.
Before I knew the word entrepreneur, I was already doing it.
The vending machine
The vending machine at school kept rejecting dollar bills. Almost everything in it cost 75 cents, and most kids only had a dollar. So I started showing up with about twenty dollars in quarters. Give me your dollar, take three quarters. Everybody was happy. I kept a quarter every time.
It worked until the school called an assembly about it.
The CD burner
Around thirteen, I got a CD burner. I'd see mix CDs advertised on TV, copy down the track lists, find the songs, burn my own versions, and sell them.
Nobody taught me either one. I saw something broken or missing and wanted to see if I could make it work. My parents gave me what I needed and more. I still wanted money that was mine, because I wanted the independence that came with it.
“If everybody's going left, I just want to know what's on the right. I want the option.”
Where it came from
Two Parents. Two Different Lessons.
My father
Work. Discipline. Try anyway.
My father taught me through action. Basketball was part of how he did it: up at 5 a.m. to train, a full school day, practice after, homework, repeat.
His other lesson is the one I took most: he will fail ten times before most people try once. I took that with me. I am willing to try something I believe in before I know whether it will work.
My mother
Education first.
My mother and her side of the family put education first, above sports and everything else. School did not come easily to me, especially standardized tests. I could work twice as hard just to get a C.
It taught me I can keep going at something that does not reward me right away.
What I'm good at
I Take Things From Idea to Real.
By eighteen, I knew I was going into business.
Three friends and I put up our own money for an event in college. We doubled our money on the first one, kept going, and turned it into a real events business.
That is where I learned one of the things I am best at: bringing different people and moving parts together around one outcome.
Events looked nothing like AI. The underlying job was more similar than it sounds. Customers, talent, partners, timing, money, logistics, people — all of it had to come together and actually work.
ZERO → ONE
I do not need to be the person who invented every piece. My strongest work is taking something from possibility to operating reality.
At a glance
Three Periods.
2006–2020
Events. At 18, I bootstrapped what became the number one college party in America, reaching about $6 million ARR while finishing Georgetown University McDonough School of Business. By 2019, the business was doing about $11M ARR across music events and talent. In 2020, COVID shut down three of my businesses overnight.
2021–2026
Tech. I angel invested, started and built multiple businesses, and spent about a year and a half finding product-market fit for Gunner.
Today
Persistent Intelligence. OA is the research lab. Gunner is the operating system we built. Okren is the commercial proof, built around the authority problem that appeared once AI action became consequential.
What I got wrong
Decisions I'd Make Differently.
I make mistakes constantly. The useful ones are the ones that change how I operate.
One
I stayed too long.
After COVID, I paid my team for a full year because I believed the business would come back. It did not come back in time.
Loyalty mattered to me, but belief in a recovery kept me committed longer than reality justified. I learned that persistence and waiting are not automatically the same thing.
Two
I handled a partnership badly.
I put about $150,000 into Repply, a CrossFit rep-counting app, and I damaged my relationship with the developer. That one is on me.
I also served as interim CEO and worked through more than 200 customer interviews. Repply went on to raise a $2 million Series A and was later sold to a competitor. Those outcomes do not change the mistake I made in the developer relationship.
I also put about $200,000 into a restaurant venture that never opened. In both cases, I moved before I had looked long enough.
Three
I go too far into one objective.
When I believe in something, my intensity can turn into control and overcommitment. Balance gets worse. I become too involved. Decisions route through me that should eventually belong to someone else.
The answer is not to become less ambitious. It is to build better structure around the ambition.
And one I got right
I walked away from Memberra.
I still liked the product. I still would have used it. I walked away because I no longer believed the opportunity matched the time horizon I wanted.
I want to be persistent about the objective, not sentimental about the vehicle.
Eighteen months. Five engineers. In the App Store. About 2,000 app members. Five paying founder-led venues. Then I stopped.
“Rushing into new projects, taking more time.”Lessons learned, Oct 2020
“Don't rush into new projects, take your time, enjoy the process/quality.”Lessons for 2021
How I work
Independent. Curious. Relentless.
Independent
I want to make my own call.
When everybody is going one way, I do not need to go the other way. I want to understand why they are going there before I decide. The vending-machine story is the simplest version of it.
Curious
I like people who are different from me.
Most of my friends are a little strange. I mean that as a compliment. The best thing someone can do for me is ask me a question I have never thought about before.
Relentless
I seek out hard problems.
How do you know what you’re capable of without trying new things? I left a field where I knew how to win and moved into one where I did not have the background, repetition, or certainty. I am here for the next two decades.
The tension in how I decide
I Broaden Before I Narrow.
I enter through movement of mind. Language, perception, interpretation, speed, wit, and mental agility shape how I explore an idea. I can see five angles before someone has finished explaining one.
I move. I pivot. I reframe. I question the question. I turn an idea around to find the contradiction inside it. I am fascinated by how thought becomes language, how thought shapes identity, and how the same reality can look different from another point of view.
That movement is how I get closer to what is true. I broaden to understand what I might be missing. Then I narrow, make the call, and move. That speed can also make me too directive, so I need people willing to challenge me when I have missed something.
Five rules I work by
How I Turn the Pattern Into Decisions.
01
Think enormously. Execute narrowly. Let reality correct the path.
Dream as large as the problem. Start with the next concrete step. Change when the world says so.
02
Don't rush into new projects.
Curiosity opens doors. Discipline decides which one deserves years.
03
Partner only with people who live in the pursuit of excellence.
I use different minds and different tools to widen what I can see. The decision is still mine.
04
Be unreasonable in all things, as long as it's for the greater good.
Ambition is only useful when it's pointed at something worth the cost.
05
Capability does not create authority.
The thing that made this concrete for me was intelligent systems operating around real money. It's why Okren exists.
Values
Nine Values, in My Words.
The nine on my plan. A few of them now have stories attached.
01
Family
The most important thing in life. It brings me the most joy: being, learning, furthering, developing.
02
Curiosity
I get this from my father. I'm curious about what I don't already know.
I deliberately seek out people who see the world differently from me.03
Honesty
Living with an open heart. No lying, no point. Having the vocabulary to say what's really on my mind.
04
Discipline
One of the most important traits I learned from my family. It guided me to where I am now.
Five a.m. basketball with my father.05
Generosity
Add value to others and it always comes back to you.
Sometimes somebody just needs you to sit there and listen.06
Productivity
One of my life's missions: to be a very productive human. I'm most happy when I am.
07
Adventurous
I want to try almost everything once. Life begins at the end of our comfort zone.
08
Positive energy
My guiding force these days. I stay away from negative energy.
09
Respect
Treat everyone the way you want to be treated. Simple.
The plan over time
The Questions Narrowed. The Values Didn't.
The questions existed before the plan. October 2020 is when I started writing them down seriously enough to compare one version of myself with the next.
These lines come from my dated planning documents. I've left out the personal material and the parts that were never meant for anyone else.
October 2020
The question appears
“Tech (efficiency): Want to create an algorithm that makes life more efficient.”
“Focus on 2030–2035 — Develop comprehensive knowledge to prepare for what the world will be.”
“What does 2030 look like, with the confluence of Artificial Intelligence, Machine Learning, Robotics…”
2023
AI becomes explicit
“Launch a tech unicorn making people’s life more productive.”
2026
The work becomes specific
“AI Native Execution Systems.”
“Teach founders how to think in decades.”
“What 2030–2040 looks like when AI compresses labor value and debt systems strain.”
Today
What survived
Persistent Intelligence is the larger question I am willing to spend the next decade on. Okren is the authority problem we found once intelligent action became consequential.
The plan did not predict the path. It helped me see which questions kept surviving it.
These excerpts came from private planning documents I've kept over time. The originals contain personal material and aren't published in full. I am showing four public snapshots here because they carry the clearest evolution; the intervening plans remain private support.
What I'm building now
Three Things, and My Job in Each.
OA
Persistent Intelligence.
OA is the research lab and long-horizon home for the larger question: what intelligence needs around the model to understand situations, make sense of change, act, observe what happened, and continue over time.
My job. Keep following the question and decide which problems deserve to become real work.Explore OA →Gunner
The operating system.
Gunner is the operating system for Persistent Intelligence. The technical architect I work with built it. I did not write the system.
My job. Understand what it makes possible, define what the applications need, and turn those possibilities into products and companies.See the architecture →Okren
The commercial proof.
Okren is the commercial proof of this work: software that helps people supervise AI-initiated financial actions before they become real.
My job. I recognized the authority problem Gunner exposed and turned that problem into a company — product direction, customers, commercialization and operating decisions.See Okren ↗Gunner operates. Okren supervises.
Why this problem
I will spend the next two decades on it.
I expect AI to become far more capable. I do not think increasing capability should automatically mean increasing authority.
Persistent Intelligence is the larger line of work.
Okren is the separate authority problem that appeared once intelligent action became consequential.
What I Am Betting On
The models will change. The work should survive them.
I believe AI will become far more capable, and the models themselves will keep changing. The valuable thing will be the objectives, memory, knowledge and responsibility that carry through those changes.
Gunner is our answer to the operating problem. Okren is our answer to the separate question of human authority over AI-initiated financial action. My bet is that both responsibilities become infrastructure as AI takes on lasting work.
I am not trying to build around one model cycle. I am trying to build what the next cycles will need.
What I'm still learning
What's Unproven.
Okren’s first paying external customer is evidence that somebody will pay for this and use it. It is not proof of a repeatable company.
The company
Repeatability.
Can a bounded first deployment expand into more workflows, more customers and repeatable institutional adoption?
The organization
Founder independence.
Can I build strong operators who own their areas so the company becomes less dependent on me being in every loop?
The thesis
Durability.
Which parts of Persistent Intelligence remain durable responsibilities, and which parts get absorbed into the broader AI stack?
Me
Intensity without bottleneck.
Can I keep the intensity that helps me build while putting enough structure around it that it stops becoming the constraint?
1paying external customer
Real moneybounded live workflow
≈1,440/daygoverned cycles in the current path
<1 mstypical evaluation in that path
Recordedgoverned decisions are reconstructable
I do not yet have a clean evidence-based stop rule for the broader thesis. Belief is not enough. The next stage needs more external evidence, not more conviction.
Health
Discipline You Can See.
Health has its own line in my plan, next to the work. I do CrossFit, I lift, and I know exactly where I stand. I like knowing that discipline exists somewhere I can measure it.
Wim Hof breathing and advanced neurofeedback are how I work on the mental side.
CrossFitStrength trainingWim Hof breathingNeurofeedback
Reading list
What's on the Shelf.
Fifteen from the reading list in my plan.
MasteryRobert Greene
BE 2.0Jim Collins & William Lazier
The Infinite GameSimon Sinek
EssentialismGreg McKeown
The AlgorithmJon McNeill
Never Split the DifferenceChris Voss & Tahl Raz
The Culture MapErin Meyer
The Algebra of WealthScott Galloway
Build the Life You WantArthur C. Brooks & Oprah Winfrey
Prisoners of GeographyTim Marshall
The Fourth TurningWilliam Strauss & Neil Howe
EducatedTara Westover
The Devil You KnowCharles M. Blow
The Way of the Superior ManDavid Deida
The Immortality KeyBrian C. Muraresku
Interests
Nine Things I Keep Thinking About.
1–3 of 10
From the interests list in my plan. Click any card to open it.
Quotes
Words I Keep.
Be unreasonable in all things, as long as it's for the greater good.
Allow, invest, and partner with only those people who live in the pursuit of excellence.
Intelligence, dreams, and ability are evenly distributed across humanity; empowerment and opportunity are not, and that's why we have the responsibility we do.
To progress again, man must remake himself. And he cannot remake himself without suffering. For he is both the marble and the sculptor. In order to uncover his true visage, he must shatter his own substance with heavy blows of his hammer.
The more tranquil a man becomes, the greater is his success, his influence, his power for good. Calmness of mind is one of the beautiful jewels of wisdom.
The challenge of leadership is to be strong, but not rude; be kind, but not weak; be bold, but not a bully; be thoughtful, but not lazy; be humble, but not timid; be proud, but not arrogant; have humor, but without folly.
When a man comes to a conscious vital realization of those great spiritual forces within himself and begins to use those forces in science, in business and in life, his progress in the future will be unparalleled.
Music
What's Playing.
The playlist is in my plan. These covers are a few of the records behind it.
The idea
I Think Everyone Should Write One.
A business plan asks what you are building, why it matters, what could go wrong, and what you will do about it. I started asking the same questions of my own life.
I wrote the first plan in October 2020. I've kept rewriting it. The values have been more stable than the answers. The questions have narrowed.
“Leave a positive legacy on this earth. Help change lives. Help family and friends reach their goals.”
From my plans, 2020 to 2026.
The tool behind the page
Eight Questions. Revisit as Life Changes.
This is the compressed version of the exercise I keep doing myself. A life business plan does not need to be impressive. It needs to be honest enough to check against reality.
- What am I building, and for whom?
- What do I value, shown in what I actually do?
- What do I believe about the next ten years?
- What did I get wrong this year, and what do I do differently now?
- Who do I need, and who needs me?
- What will I stop doing?
- What would make me change course?
- What does a good day look like in five years?
Site Updated October 17th, 2026
Privacy
Privacy.
What this site collects, and why.
How OA handles information on this website
01
Ask OA.
When you ask Ask OA a question, the site sends your message, a limited amount of the preceding chat, and OA’s public Context File to OA’s Vercel function and Google’s Gemini API to generate an answer. OA does not maintain its own database of chat transcripts. The current chat remains visible in your browser tab until you clear it or leave. On Google’s free Gemini API tier, Google may use submitted questions and generated responses to improve its products and services, and human reviewers may examine them. Do not submit confidential, sensitive, or personal information. If OA later switches to paid API processing, these terms and this notice will be updated. OA does not sell your messages.
02
Sensitive information.
Ask OA is a public information service. Please do not submit sensitive personal or confidential information. Conversations are not covered by a confidentiality agreement. For sensitive discussions, contact Freddie directly.
03
AI answers.
Ask OA answers may be incorrect. They are informational, not professional advice or commitments by OA.
04
Hosting and technical data.
Vercel handles technical request information, such as IP addresses and browser information, when serving the website. Hosting and abuse-prevention systems may retain technical logs. Some speaker photographs, video thumbnails, and book covers are loaded from third-party image hosts. Those hosts may receive your IP address and browser information when the images load. Selecting an external source opens that provider’s website.
05
Contact OA directly.
You can contact Freddie by WhatsApp or email. The WhatsApp link opens a conversation with +31 6 18 84 24 72; WhatsApp handles messages under its own terms. Email links open a message to freddie@oaagency.com. Messages you choose to send are handled by those services and received by Freddie to respond.
06
Information requests.
Email freddie@oaagency.com to request access to, correction of, or deletion of information OA holds about you. Email, WhatsApp, and other third-party services have their own data-handling practices.
07
Changes.
OA updates this notice as its services change. The current site does not add advertising or analytics tracking scripts to the HTML.