Reflection agent

Reflection is AutoZyme’s feedback loop. After a real run, it captures what was confusing, missing, redundant, or awkward, then sends that feedback back to the framework.

The point is self-evolution: the more people report sharp, concrete friction from real tasks, the better the prompts, CLI checks, docs, and autotests become for the next run.

Copy the prompt

101 lines · 6,237 chars

Paste it into your coding agent from your task directory after finishing a phase. The agent asks which phase you just finished, writes the feedback files, and opens a small PR to the AutoZyme repo.

When to run it

Run it after any completed phase:

Finished phasePick category
Benchmark initializationinitialization
Optimization loopiteration
Validationscaling
Packagingpackaging

Total human time is usually a few minutes.

What matters

One concrete observation is enough. Good reflection feedback looks like:

  • a prompt step was unclear or in the wrong order
  • a zyme command had confusing output
  • the agent got stuck because an expected file or check was missing
  • a validation or audit rule should become an autotest
  • docs failed to explain something you needed during the run

The agent writes prompt feedback and CLI feedback separately. Empty is fine; if you only have one useful note, send that one.

What happens next

Maintainers batch-review reflection PRs. Repeated patterns become prompt edits, CLI improvements, docs changes, or new tests. Over time, real user friction becomes framework behavior instead of tribal knowledge.