Let Your Agent Finish the Job

An agent can be bad at the first draft and still be good at the task. What matters is whether it has a way to improve before it brings the work back.

The best agents we’ve built came from looking at how people do good work. A person can hand a draft to someone else, get feedback, and take another pass.

Coding agents can run tests, find a failure, and keep working. They can make several attempts without needing a person to inspect each one. Tests give them feedback beyond their own view that the code is done.

That loop is much harder to build for a presentation or an interface. An agent can hand you a draft that misses basic requirements, then wait for you to supply every correction.

Set the bar before you review

The place to start is a set of ground rules the work must meet before it comes back to you.

For an interface, that might mean the page is easy to understand, has one clear purpose, and uses icons to explain something rather than fill space. Those rules will not decide every detail. They give the agent specific problems to check and fix.

Firmament’s review gates provide that feedback. The agent can submit a draft, get told where it falls short, revise it, and submit again against the same rules.

You can then ask it to come back when the work passes those checks. It is ready for your review at that point. You still judge the result, but the agent has room to make more of its own progress before it asks for your time.