The agent you can ignore

The first month with a new agent, you watch it like a hawk. Every run, you open the log. Every output, you double-check against what you would have done. That’s not paranoia, that’s correct. A new hire gets the same treatment. The interesting question is what happens in month three.

If you’re still checking every run in month three, the agent failed. Not loudly. It probably produces fine work most of the time. But it never crossed the line that actually matters: it never became something you could stop thinking about. And an agent you can’t stop thinking about isn’t saving you the work. It’s moving the work from doing to supervising, which for most knowledge tasks costs about the same.

We talk about agents in the language of capability. Can it write the report, close the books, triage the inbox. Capability is table stakes and it’s also the easy part. The hard part, the part that decides whether the thing is worth having, is whether it earns the right to be ignored.

Boring is the milestone

Think about the software you actually trust. Your payroll runs. Your DNS resolves. Your backups complete. You don’t watch any of it, and you’d struggle to remember the last time you thought about it at all. That forgetting is not neglect. It’s the highest compliment infrastructure can earn, and it was earned the slow way: by being correct so many times in a row that checking became a waste of your attention.

An agent is on the same path, and the path is unglamorous. It does not run through a better demo. A demo proves the agent can do the task once, under good conditions, while someone watches. Ignorability is the opposite claim. It says the agent does the task on the bad days too, when the input is malformed, the upstream API is slow, the data is half-missing, and nobody is looking. You can’t demo that. You can only accumulate it.

This is why the right metric isn’t accuracy on a good run. It’s how long the agent goes between surprises. A surprise is any moment it forces you back into the loop: a silent failure, a confident wrong answer, an edge case it handled by guessing. Every surprise resets the clock and pulls you back to watching. A long quiet stretch is what lets you finally look away.

Which means the work of building a trustworthy agent is mostly the unsexy work. Handle the malformed input. Fail loudly instead of quietly. Say “I’m not sure” instead of inventing. Survive the restart. Leave a trace you can audit later so you don’t have to watch in real time. None of it shows up in a highlight reel. All of it shows up in whether you sleep.

There’s a tempting counterargument: maybe you should always keep a human in the loop, just to be safe. For genuinely high-stakes, irreversible actions, yes, and a good agent asks for that gate itself. But “watch everything forever” isn’t caution, it’s a confession that the agent never got reliable enough to trust, and a permanent human reviewer is the most expensive feature you can ship.

So the bar we hold ourselves to is deliberately anticlimactic. Not “look what it can do.” The real test comes a month later, when you realize you stopped opening the log and nothing broke. The day an agent becomes boring is the day it starts paying for itself. We’re trying to build the most boring coworker you’ve ever had.