Recent work - https://arxiv.org/pdf/2601.03768 - https://arxiv.org/pdf/2602.20082 (idea: running multiple agents to solve proof goals in parallel)

  • Tips:
    • Which model: Use the best model
      • it is always more money-saving to get your code right at the first time
    • Which Agent/CLI:
      • Not OpenClaw: waste of tokens and risk of hallucination
    • Use MCP:
      • Rocq: vsrocq PR #1194 todo check if this is the one I am using
      • Lean: lean4-mcp
  • Caveats:
    • Agent can cheat:
      • false admits
      • silent definitions changes
      • original lemma changed
    • Put hard rules in AGENTS.md, CLAUDE.md
  • Setup Remote Access to a persistent running agent
  • todo Git Worktree to Run Multiple Agents on the same project (check OpenAI’s official Codex documentation)
  • Use voice inputs: Use typeless or similar voice to text tools.