Two runs on one scratch project, deterministic, no key and no network:
dotnet run --project samples/AgentExperience.Sample.Reflective
dotnet run --project samples/AgentExperience.Sample.Reflective -- --compareSample A, a reflective coding agent. The first attempt applies bump-version, the build fails, the evaluator
proposes "pin the dependency", the retry applies pin-version, and the host's own build check passes. Both attempts are
one captured run, finalized once, stored as one verified lesson.
Sample B, reflection plus experience. A related task starts with the lesson Sample A stored, applies the right fix on its first attempt, and the reflection loop never has to retry.
--compare runs four arms (neither, reflection only, memory only, both). It shows the mechanics, not a benefit: the
scripted model does whatever it is handed, so the numbers say nothing about a real model. Whether reflection or memory
helps your tasks is what the injection holdout and a real model run are for; no improvement is claimed here.
To try a real model, replace ScriptedModel in Program.cs with any IChatClient. The loop, the build check and the
stored lessons work the same way. Read Reflective execution first.