An agent earns the next job by showing the receipt

Two agents do the same job equally well. One returns an answer. The other returns the answer plus the three records it read, the rule it applied, the one case it flagged as uncertain, and the tool call it made to check. You will give the second agent harder work next week. Not because it is smarter. It is not. You will give it harder work because you can audit it, and you cannot audit the first one at all.

This is the quiet mechanism behind every real deployment of autonomous work. People talk about trust as if it were a feeling, a vibe the agent earns by being impressive. It is not a feeling. It is a function of evidence. An agent gets more responsibility exactly as fast as it can show its work, and not one rung faster.

Capability is invisible, and invisible cannot be promoted

Think about why a human gets promoted. It is almost never raw talent, because talent is unobservable. What gets observed is the trail: the well-reasoned recommendation, the caught error, the decision you could follow and agree with. A brilliant employee who hands you only conclusions and never the path to them stays exactly where they are, because you have no way to know whether the next conclusion will be the one that quietly costs you a customer. The thing that moves them up is legibility. You can see how they think, so you can extend how far they reach.

Agents are the same, only the stakes are sharper, because an agent that hides its work hides it perfectly. A human at least looks nervous when guessing. An agent states a fabricated total in the same confident sentence it uses for a true one. So the only way an agent earns the next job is by making its work cheap to check. The receipt is the product. The answer is just the part that happens to be quotable.

A receipt is concrete, not decorative. It is the records the agent actually read, not a summary of what it might have. It is the tool it called and what came back. It is the rule it applied and the threshold where it chose to stop and ask. It is the explicit I am not sure about this one, surfaced before it matters rather than discovered after. None of that makes the agent better at the task. All of it makes the agent safe to extend, which is a different and more valuable thing. A capable agent you cannot see is a liability with good output. A modest agent you can fully audit is a teammate you can grow.

This inverts how most teams roll out automation. They wait for the agent to be good enough to trust blindly, which is a bar it will never clear, because blind trust is not a state any sane operator reaches. The agents that actually take on more are not the ones that got more capable in the dark. They are the ones that got more transparent, so that handing them the next responsibility stopped being a leap of faith and became a reading of the receipts. You expand the agent’s job because the trail says you can, and you can always read the trail.

So the feature to build first is not a smarter model. It is the receipt. Make the agent narrate what it read, what it called, what it was unsure of, every time, even when it is right, especially when it is right, because the receipt on the easy cases is what lets you believe it on the hard one. An agent that cannot show its work is stuck at its first job forever, no matter how good it is, because nobody can responsibly give it a second. An agent that shows the receipt climbs, because every job it does leaves behind the exact evidence that justifies the next.

Trust is not granted to the capable. It is granted to the legible. Build the legible agent.