Closing the loop: agents that verify their own work

Closing the loop is what separates an autonomous agent from a one-and-done tool. A one-shot tool generates code and hands it to you. An agent that closes the loop runs the tests, reads the failure, fixes its own work and tries again, without waiting for a person to notice something broke.

I treat this as an architecture principle rather than a feature. In a harness, these feedback mechanisms are the sensors: compilers, linters, unit tests, policy checks and schema validators that reject a change and force the agent to correct itself before a human ever looks. The model supplies the effort. The sensors supply the judgment you can actually trust.

The failure mode I see most often is teams investing in a better model and leaving the sensors weak. If your test suite is thin, the agent will happily converge on code that passes thin tests. Self-correction is only as good as the signal it corrects against, so the quality of your validators sets the ceiling on how much autonomy you can safely grant.

The practical rule: before you widen an agent's permissions, strengthen what checks its work.