Let Copilot Agent Mode handle coding, then constrain it

A developer let Copilot Agent Mode do all the coding, found they understood their code less, and moved to using it more narrowly.

What They Did

The poster initially used GitHub Copilot Agent Mode for nearly everything: tests, refactors, and new modules. After finding that blind delegation made them understand and debug their own code less, they changed to a tighter loop where the agent still proposes work but they interrogate its approach against their architecture and review the output instead of just accepting it.

How It Went

The reported result was qualitative: they felt the shift from “do everything” to “challenge me” made them a better developer and helped them keep ownership of the codebase. In comments, others said success depended heavily on a well-structured repo, and one claimed full features worked about 90% of the time.

Worth Knowing

The main catch was loss of understanding: when the agent handled too much, debugging got harder and the developer felt they were slipping into autopilot.