03 · Improve

Mutation Agent

Implement the change, validated, before it ships

Mutation Agent covers the implementation half of the Optimize phase. It is an instruction set handed to a coding agent specialised on AI-agent engineering. The coding agent executes the change exactly as Optimize Agent specified: mutate the prompt, add the tool, update the dataset row, swap the model, restructure the loop. The result runs through the Evaluator framework before shipping via your target platform, whether that is a Langfuse prompt version, a GitHub PR, or a pipeline queue.

Mutation Agent does not decide what to change, that is Optimize Agent's job. It does not diagnose, that is Diagnose Agent's job. It does not invent strategies on its own. Its discipline is execution with validation: implement the specified mutation, validate it, ship the candidates that beat baseline, abandon the ones that do not.

What it does

  • Coding-agent instruction set per Optimize strategy
  • Concrete change applied (prompt diff, tool addition, dataset row, model swap, config update)
  • Pre-merge validation via the Evaluator framework
  • Auto-application via your target platform (Langfuse prompt version, GitHub PR, pipeline queue)
  • Audit trail per shipped change with a rollback hook

Inputs

  • Optimize Agent strategy
  • Coding-agent runtime
  • Target platform credentials
  • Evaluator framework hook

Outputs

  • Validated mutation
  • Pull request or platform-native artifact
  • Audit trail with rollback hook

Works with

GitHubGitHub
GitLabGitLab
LangfuseLangfuse
ClaudeClaude
CursorCursor

Try Mutation Agent today

Install the CLI and run this agent against your own evals in under five minutes.

See a mutation run