Briefing
The article casts the agent as the coder and the author as the evaluator, who checks its work against a written product spec, cargo check, UART traces and the bench. Prompts stay small, and when debugging stalls he hands the agent a hypothesis to test. In one case that showed the device resetting before a communication window finished.
His recommendations: pin crate versions, keep known-good ESP-IDF C examples in the workspace as a reference for protocols such as BLE provisioning, and commit and tag each milestone that works on the bench. His failure modes include invented APIs, over-refactoring and a debug journal the agent deleted when asked to clean out garbage files.
Written by the Boards & Bots curator
“Speed without tracing, commits, and occasional tidy-up is how you lose afternoons to regressions nobody can bisect.”