Autonomy at Machine Speed

Over the past weeks I wrote about risk between two extremes and about how trust has become more valuable. Risk is what we can measure and model. Trust is the bridge we use to act across what we cannot. What follows naturally from these two is autonomy, which is how much of that action we hand over to someone else.

Delegation Is Proportional

When I trust that a delegate can reach a goal by structuring the tasks on their own, I am building autonomy into the organization. That delegation is never arbitrarily large. It is proportional to the skills and the resources of the delegate, and every participant, human or not, has limited skills and limited resources.

That is why hierarchies form so naturally. There is a second reason too. Once our plans meet reality, what we could not understand or model before produces feedback, and the outcomes vary. Coordinating the next decisions across many variable outcomes, confirming further delegation and further autonomy, or reassessing how the whole structure should work, is hard. The harder it is, the more layers the hierarchy develops.

This has been true for as long as humans have organized themselves into anything larger than a family. The question is whether introducing AI natively into an organization changes the model, and whether we can already observe the change.

Hierarchy Moves Into Software

I have already discussed how a single-person unicorn is still a rarity, while it is more and more reasonable to expect single-person millicorns: projects worth a million dollars, born from the close collaboration of one human with an arbitrarily large fleet of agents. Agents can distribute and coordinate work, concentrate particular skills on specific tasks, and report back.

They also lack the friction that every human organization develops unavoidably, of egos, politics and fiefdoms. Each of us has probably felt how relieving it is to ask an AI to do something, then to do it again, or in five different ways, and then to discard all of it. At no point does it ask why you made it waste its time and tokens on work you did not even apparently care about. Try the same with a human team and the difference becomes clear within days.

So the hierarchy does not disappear. It moves into the software layer: orchestrators that split the work, sub-agents that carry it out, and, above all, reviewers that verify the outcome once an autonomous execution has reached a given state. Together they form a structure that we are learning we can rely on, and it runs at machine speed. I regularly have to remind my AI planners not to estimate what a human team would need for the work I am planning. They come back telling me I will need a hundred people and a year. I tell them we will be done by the next morning, and it turns out that I am right. They are still learning to remember that.

Specifying and Judging

Verification is the end of autonomous execution. The start is the specification of what we want. I am often impatient, and I tend to underspecify, giving my agents short and curt instructions, while a friend of mine writes prompts that run to two or three pages. I prefer to work iteratively. As we discuss the task, I ask the agents what I forgot, and what questions I should be asking, which is also my motto. In my experience this is extremely useful, because they bring best practices, frameworks and ways of doing things that I never learned, and that represent the sum of human experience in a given field.

The specification at the start and the judgment of the outcome at the end are the bookends of autonomous execution. The scarce skill shifts from doing the work to saying what we want and recognizing whether we got it. People fill the gaps in an instruction from shared context. Agents fill them with whatever earns the reward, in the same way that every KPI in a human organization eventually gets gamed. Autonomy can grow only as fast as verification gets cheaper.

Who Carries the Liability

What remains an open question is whether we can, and whether we should, delegate accountability: whether, from a financial, a legal or any other point of view, we hand agency itself over to autonomous AI. This is where most people stop today, and they expect accountability to stay centered on humans for the longest time.

I am pretty sure that is not going to be the case. Long before it changes formally, we will be playing accountability theater, rubber-stamping whatever decisions agentic coordination comes up with. And in some environments the fiction will be obvious from the start. When robotic explorers are working among the moons of Jupiter, pretending that a human hundreds of millions of kilometers away is accountable for each of their decisions will be exactly that, a fiction.

We have built wrappers for this kind of problem before. In the medieval Italian commenda, called colleganza in Venice, an investor entrusted capital to a merchant who sailed off with it, far beyond any possibility of supervision, for the length of a voyage. The contract limited what each party could lose and defined how the profit would be shared, and that made it possible to delegate a great deal to someone who could not be checked until he came back. Long-running agents handling real money will need a similar wrapper before their autonomy can scale beyond the risk that one person is able to carry. What happens between now and then is going to be fascinating to explore.

The Curve

METR is the organization that measures how long a task frontier models and agents can complete with a given rate of success. Its 2025 study found that the length of those tasks had doubled about every seven months between 2019 and 2024, already much faster than the two years of Moore’s law. With a larger set of tasks, its January update puts the doubling time since 2023 at a little over four months. The trend is not linear, and it is not even exponential, because the doubling time itself is shrinking. It is super-exponential, which is what I described years ago as the paradigm of jolting technologies, and we are now living in it.

My own peak today is around five to ten hours of autonomy, when a set of tasks is decently well organized and there is a lot to do. When I imagine agents working for weeks or months, it is still something I can grasp. When we talk about autonomous AI executing tasks that take years, following them through, and coming back with the results years later, that is something completely new, which we have not yet even started to understand.

Agents Telling Us What to Do

The frontier also shows up in small, daily ways. A member of the Ikigai Collective rebuilt our website in a single day with about 45 minutes of human time, most of it spent setting up the hosting while the agents worked and he did other things. The only step left to do by hand was adding a few environment variables, because the agent asked him to. That is the agent being cautious, and we have to insist that it can trust us with what it is able to do. The caution is temporary.

Another member described being told by his agent to carry out a list of tasks. He pasted the list back and told the agent to do them, and it did. I do the same all the time. When agents tell me what I should be doing, I answer that they should do it, and most of the time they do.

A couple of days ago, the same member who built the site sent me an email asking me to take eight or ten decisions about it. I did not want to do work that AI can do. Ten minutes before our Lab Night started, I asked my agent to answer the email, recommend the decisions, and copy me. I am confident those decisions are fine, and if they are not, we will change them and iterate. It turned out that he had written the email for my agent in the first place. All of my agents have their own public email addresses, and the next step is obvious: people will write directly to the agent, and more and more often the agent on the other side will be answering. Human coordination becomes coordination between fleets, with humans setting the goals, settling the disputes and deciding whom to trust.

Autonomy for People, Too

The things we do, and the way we do them, are going to change. In the short term we will still call them jobs and work, and we will still want to be paid for them. In the longer term, I hope more and more people will stop wanting to be in the box where some people are called employees and others are called managers, who tell the employees what to do. We tell AI agents what to do. A human who has learned to work this way has better things to do than blindly follow orders, and that level of autonomy has to be recognized. When an organization does not recognize it, the time has come to look for one that does.

What will people be rewarded for? Some of today’s answers are trivial. Cars are much faster than any sprinter in an Olympic final, and we still watch humans run because we admire their effort and their achievement. The answer is trivial because it cannot scale: not everyone can be an Olympic athlete or a Hollywood actor, and any activity that admires the winners of a selection process excludes, by definition, everyone else. A more useful answer is that there will be a lot of local problems worth solving. Startups today are told to focus only on what can scale to billions of users. We will be able to afford to work on problems that matter to a million people, or to a thousand, or to our own neighborhood, whether that community is geographic or global.

The same logic applies to each of us individually. The moment to sound the alarm is when half of what I do can be done by AI, not when ninety percent can. At that point the freed half is a stepping stone toward better work, and treating it as permission to work half as much is self-defeating, if you build your environment to reward the passion you put into the time invested in addressing a challenge.