moai-workflow-research

Automate workflow research experiments by running iterative binary-eval tests that compare new components against proven baselines. The system organizes eval suites and changelogs, applying your evaluation schema to systematically identify optimization opportunities.

A binary-eval experimentation loop iteratively tests new skills, agents, rules, or configs against a baseline to measure pass or fail outcomes. It organizes observations across isolation worktrees to validate component changes before integration.

Ensure safety during experimentation by using a five-layer architecture that includes FrozenGuard, worktree sandboxes, canary regressions, rate limiting, and human approval. These safeguards prevent unsafe modifications from affecting baseline workflows.

Yes, you need a defined evaluation schema specifying runs_per_experiment, max_experiments, and a pass/fail threshold. You must also provide data locations for eval suites, baselines, experiments, and observations to execute the research workflow.

Organize isolated research cycles by separating eval suites, baselines, experiments, and changelogs into distinct worktrees. This isolation lets you test Moai-ADK components systematically and review results on dashboards without risking the baseline.