Steering Agents To Use Feedback Loops With Skills.mp4
Feedback loops are how you keep agents on the right path. Steering with AGENTS.md files and skills can make an agent more likely to do the right thing, but feedback loops give the agent concrete signals about whether its work is actually correct.
The agent writes code, receives feedback from the codebase, adjusts, and repeats until the result is good enough.
Great engineers do not rely on instinct alone. They use feedback loops to check their own work:
The same idea applies to AI agents. If you want better code from an agent, give it better feedback.
Without feedback loops, the agent can produce code that looks plausible but breaks real behavior. With feedback loops, the agent has a way to test its assumptions against the codebase.
A strongly typed language like TypeScript gives the agent immediate feedback when it uses the wrong shape, calls an API incorrectly, or introduces a typo. That feedback helps the agent correct itself before the bug reaches a user.
Automated tests give even stronger feedback. Tests run the code and check whether the behavior matches expectations. A good test suite gives the agent a repeatable way to ask, "Did this change work?"
The better your types and tests are, the better the signal you can give the agent.
This course repo already has feedback loops available.