Earned Autonomy
Autonomy should not be a dial you set on day one. It is authority an agent earns by proving, decision after decision, that it can be trusted.
Founder, Task Machine
Most tools that let an agent act on its own also ask you to decide, up front, how far it can go. You pick a level. Low, and every action waits for your approval. High, and the agent proceeds alone. You make this choice on the agent's first day, before it has done a single thing you can judge.
Consider what that choice actually is. You are setting how much to trust something with no track record. Set the level low and you have hired a capable worker and then insisted on watching its hands. Set it high and you have handed real authority to something you have no reason to trust yet. Trust granted before evidence is not trust. It is a guess wearing trust's clothes.
Authority follows a record
People do not grant authority this way. A new hire does not get signing power on the first morning. They get a small mandate, they use it, and the mandate grows as their judgment proves sound. Authority follows a record. That is not caution for its own sake. It is how you delegate to someone whose reliability you cannot yet see.
Economists have a name for the underlying difficulty. The principal-agent problem describes what happens whenever one party delegates work to another: the agent may not act in the principal's interest, and checking every action is expensive. Michael Jensen and William Meckling built much of the modern theory of the firm on this tension, and Oliver Williamson spent a career studying the machinery organizations use to manage it. Most of the classic answers concern incentives and monitoring. There is a third lever that matters more for agents: calibrated trust, granted in proportion to demonstrated reliability and adjusted as that reliability changes.
Autonomy should be earned, not configured. It is not a setting. It is an authority that expands as an agent proves it can be trusted at its current level, and contracts when it cannot.
Reliability has to be measured, and measured honestly
For autonomy to be earned rather than felt, reliability has to be measured, and the measurement has to be honest. This is harder than it sounds. Suppose an agent has made a hundred decisions at its current level and you have corrected it twice. A raw success rate says ninety-eight percent, which sounds like plenty. But a hundred decisions is a small sample, and a small sample can flatter. Ten good decisions in a row is not evidence of trustworthiness. It is evidence of ten good decisions.
There is a well-worn statistical tool for exactly this. In 1927 the statistician Edwin Wilson described a way to put a confidence bound on a proportion that accounts for how much evidence you actually have. Applied here, it answers a better question than "what is the success rate." It answers "given how much this agent has actually done, what is the lowest its true reliability is likely to be." That bound is deliberately unimpressed by small samples. It grants confidence slowly, as evidence accumulates, and withholds it when the record is thin. Autonomy that expands only when the lower bound clears a high bar is autonomy that has genuinely been earned.
The other half is asymmetry. Trust should be slow to grant and quick to withdraw. A handful of failures at a level of authority should pull that authority back well before a matching handful of successes would have extended it. A system that promotes cautiously and demotes readily is one that treats your exposure as real.
Two models side by side
The difference stops being subtle once you lay the two approaches next to each other.
| Configured autonomy | Earned autonomy | |
|---|---|---|
| How it is set | A level you choose up front | An authority that accrues from a record |
| What it rests on | Your guess about a new agent | Measured reliability at the current level |
| On a run of success | Unchanged until you change it | Expands once the evidence is strong enough |
| On failure | Unchanged until you notice and lower it | Contracts on its own, and quickly |
| A thin track record | Treated the same as a long one | Treated with suspicion, as it should be |
| What it asks of you | Constant re-evaluation | Set the standard once, review the evidence |
Configured autonomy makes you the monitor. You have to notice that an agent has earned more freedom, or lost the right to what it had, and act on it. Earned autonomy makes the record do that work and brings you the moment a change is warranted, rather than asking you to catch it.
What it requires
Earned autonomy has a real prerequisite: a clear standard for what counts as a good decision. If you cannot say what success and failure look like, you cannot measure reliability, and the model quietly collapses back into guessing. It also depends on volume. For rare, high-stakes decisions there may never be enough evidence to earn much freedom, and those should stay gated no matter how well the routine work goes. The model rewards repetition. It does not manufacture confidence where there is no track record to draw on.
Setting a standard instead of a slider
This is the principle behind how Task Machine handles autonomy. Every decision that comes to you for approval is recorded against the agent that made it, at the authority level it currently holds. A statistical bound, not a raw average, decides when the evidence justifies more authority, and it demotes faster than it promotes. You can see an agent's recent approval rate, how many decisions stand behind it, and what would justify the next step. You are not dragging a slider and hoping. You are setting a standard and watching authority accumulate, or drain, on the evidence.
Earned autonomy is also what lets the inbox go quiet over time, as more of the work proves it can resolve itself. Continue with Management by Exception.