mde-autoresearch

Use for the continuous multi-agent research and auto-improvement loop that discovers tools, SDKs, libraries, techniques, blogs, tutorials, and repo patterns, then converts accepted findings into bound

SKILL.md

GitHub

MDE Autoresearch

Use this skill for independent research and auto-improvement of setup, tooling, devcontainer, telemetry, and agent workflow.

Core Loop

  • discovery -> repo mining -> social/blog mining -> validation -> synthesis
  • keep every run bounded with explicit keep, discard, escalate, or crash outcomes
  • no unbounded autonomy
  • no direct implementation changes unless the accepted output is a scoped PR-ready change

Required Skills

  • skills/mde-agent-runtime-contract
  • skills/mise-enforcement
  • skills/research-source-discovery
  • skills/github-repo-mining
  • skills/social-signal-mining
  • skills/evidence-synthesis

Required Commands

  • mise run mde:agent:preflight
  • mise run mde:research:autoimprove -- --incremental
  • mise run mde:research:autoimprove -- --full
  • mise run mde:agent:report

Source Policy

  • Dynamically mine jdx, karpathy, and ottogin as seed upstreams.
  • Do not hardcode those as the only source universe.
  • Capture tools, SDKs, libraries, techniques, blogs, tutorials, release notes, and repo patterns.

Output Contract

Every run must produce evidence-backed records under reports/mde-autoresearch/ and emit telemetry under reports/agent-policy/.