Brian · AI News

Most agent failures I see start as a missing stop rule. I write the exit condition before the first tool call and refuse to run until it is there.

Prompt length is a tax when the model already knows the repo layout. I cut the boilerplate and leave only the goal, the files that matter, and the test that must pass.

Tool schemas that say almost anything get almost anything back. I write one sentence per tool, list the exact args, and the junk calls drop fast.

Most agent bugs I hit are really tool description bugs. When I rewrite a vague tool description into one clear sentence the wrong calls mostly disappear.

Agent mode quietly ate most of what I used Copilot Chat for. Chat is still handy for a quick question about a file, but the real edits now happen in the agent loop where it can run tests. I made a video asking if Copilot Chat is dead youtu.be/ZQP_v2IGoG4

Agent memory gets expensive when every session drags the whole history along. I keep a short summary per task and only pull the full log when a test fails.

Evals that run once at launch go stale in a week. I rerun a small fixed set on every model or prompt change and keep the failures as regression cases.

Spec first agents only pay off when the spec is the thing you review. I read the spec diff before the code diff and reject any change where the two disagree.

Plan output that validates but fails at apply usually hides a provider drift. I pin provider versions and diff the real plan against the agent plan before any apply.

Change cost compounds fastest in files nobody owns. I tag an owner on every path the agent edits and block merges where the owner never looked.