For a while, when I pictured a more autonomous company, I pictured an agent that kept going.
It would wake up, inspect the world, decide what mattered, do the work, notice what changed and carry on. The obvious measure of progress seemed to be how long it could operate without me.
I think that was the wrong unit.
We have just merged the first end-to-end version of our company loop in Farstax and Agent Bridge. The interesting part is not that an agent can continue. Capable agents are already quite good at finding another useful-looking thing to do.
The part that changed my mind is where we put the boundary.
The unit of autonomy we are testing is not the agent, the task or the tool call.
It is the episode.
I kept putting the human in the wrong place
Human supervision is usually described as 'human in the loop'. That phrase hides an important question: which loop?
If I approve every action, the agent is not very autonomous. I am still managing the work, just through a chat window.
If I approve a broad goal once and let the system keep waking itself forever, I have solved the interruption problem by creating a different one. The company can keep acting after the assumptions behind the original instruction have gone stale.
We ended up between those two.
I approve a meaningful outcome and one bounded period of responsibility. During that episode, the agent owns the routine decisions. It can investigate, choose the next useful action, use its tools and adapt to what it finds without waiting for me after every step.
Then the episode ends.
What happens next is a new decision, not an automatic continuation.
That distinction sounds small. It changed the shape of the system.
An episode starts with the world as it is now
A company goal can live for weeks. An agent's useful working context should not.
So we separated the durable company state from the briefing the agent receives.
The durable part is deliberately boring: the mission, the outcome we are trying to move, the evidence we trust, and the record of what happened before. That state survives whether an agent is running or not.
Immediately before a new episode, Farstax creates a fresh briefing. It includes the current outcome, the evidence that matters now, the conclusion of the previous episode, what changed since, and a short set of operating principles.
The briefing is disposable. The evidence is not.
This matters because accumulated context has a seductive quality. If an agent has seen everything, surely it can make a better decision.
Sometimes. It can also become trapped inside yesterday's story.
A fresh episode gives us a chance to re-read reality rather than merely continue the narrative.
Inside the episode, the agent owns the work
This is the part I care about most.
We are not trying to automate my task list. We are trying to find out whether an agent can make useful decisions about what matters next.
That only works if I leave some real space for it to decide.
The episode starts with an outcome, current evidence and clear authority. It does not start with a hidden plan that I expect the agent to rediscover.
Farstax also does not learn how to run Claude or Codex itself. It hands the episode to Agent Bridge, the part that already knows how to run, record and stop agent work.
That was a useful constraint. Every time we were tempted to add another scheduler, another status model or a second way to run the work, we asked whether the existing owner could do the job.
Most of the time, it could.
The company layer became smaller as the experiment became more capable.
The agent gets to explain itself. It does not get to mark its own homework
At the end of an episode, the agent leaves behind a post-mortem and a challenged review of what happened.
That record matters. I want to know what it tried, what it learnt, where it was wrong and what it thinks should happen next.
But none of that is the business result.
If the agent rewrites the landing page and gives a convincing account of why the new version is better, the company outcome has not necessarily moved. If it fixes onboarding but nobody suitable reaches a useful result, the evidence can still say zero.
This is one of the boundaries I now think matters most in autonomous work.
The agent can narrate its actions. Reality decides whether the outcome changed.
In our current experiment, only authoritative evidence moves the outcome. A persuasive end-of-episode report cannot promote itself into success.
That makes the system less flattering.
Good.
The review is part of the work
I used to think of a post-mortem as something you did after the interesting part.
In this loop, it is part of the interesting part.
An episode is not complete merely because the agent ran out of tasks, produced code or reached the end of its budget. It has to compare what happened with the evidence and leave the next decision in a state someone else can inspect.
For now, that someone is me.
The company pauses. I can inspect the result and approve exactly one next episode, or stop.
That pause is not a failure of autonomy. It is where responsibility changes hands.
Why not just keep the agent alive?
Because companies are continuous and attention is not.
A persistent outcome may need months. The useful action at 9am may be wrong by 3pm. A customer can reply. A release can fail. A metric can move. A permission can become necessary. Something that looked like the bottleneck can turn out not to be one.
An always-on agent can react to all of that, but continuity has a cost: it becomes harder to say which assumptions still govern the work and which instruction authorised the next action.
Episodes give us a clean renewal point.
Do we still want the same outcome? Is the evidence still trustworthy? Did the previous attempt move anything? Is the proposed next step still inside the authority we intended to grant?
Then, if the answer is yes, another episode can begin.
This is less theatrical than an AI chief executive running continuously in the background. I think it may be more useful.
We did less than I expected
There is another lesson in what we did not build.
We did not create a separate way to run a 'company agent'. We did not add a hierarchy of AI managers. We did not add a company scheduler that owns the agents underneath it.
Farstax owns the long-lived company facts and the gate. Agent Bridge owns the agent work. Each side does the job it already understands.
That lets us ask a cleaner question.
Can better context, evidence and authority turn an existing capable agent into something that behaves more like a useful company participant?
If the answer is no, we will have evidence about what is missing. Perhaps specialist roles matter. Perhaps the review needs to be independent. Perhaps some classes of work can earn automatic continuation while others should always come back to the owner.
But I would rather let those needs appear in the experiment than build the org chart first.
This is not the final shape
I do not know the right size for an episode.
Too small and I become the scheduler. Too large and the boundary becomes ceremonial.
I do not know whether owner approval between every episode will survive. It may turn out to be training wheels. We may discover repeatable classes of work where the system can safely authorise its own next episode within a tighter budget.
I also do not know whether one general agent will remain enough once the work becomes genuinely cross-functional.
Those are now testable questions.
That is the bit I like.
A month ago, 'how autonomous should the company be?' could turn into an architecture discussion very quickly. Now we can ask something much more concrete: what happened during the episode, what changed in the world, and did the system earn another one?
For the moment, that is our model.
Do not make the agent immortal.
Give it responsibility. Give it evidence. Give it room to act.
Then make it earn the next episode.