agent-pool

Agent Pool

Spawn subagents from this session to run an implementer/critic/validator feedback loop. No extra terminals or setup needed.

Tip: You can run /remote-control to get a URL for monitoring this session from another device.

When Triggered

Ask the user:

• Goal: what to build or improve (be specific about scope/files)
• Max iterations? (default 5, higher for complex tasks)
• Done condition: when should the critic approve? (default: "no critical issues remain")
• Test environment (if needed): Detect project type first (check for package.json, requirements.txt, pyproject.toml, Cargo.toml, etc.). Then ask the user: "Detected [type] project. Do you need a specific env? (e.g., Python env path like D:\Anaconda\envs\microgrid, or press enter to skip)"
• Test command (optional): default by project type. User can override.

If the user already specified exact files/scope, skip the Explorer step.

The Loop

Iteration 1

1. Explorer (skip if scope is already clear) — spawn Explore subagent: "Explore [scope]. Return: relevant files, current state, key patterns."
2. Implementer — spawn general-purpose subagent: "Goal: [goal]. Context: [exploration summary]. Make the changes. Return a 1-2 line summary of what you changed."
3. Critic — spawn general-purpose subagent: "Review changes toward: [goal]. Verify by reading the actual changed files. If this is a git repo, run git diff to see exactly what changed. Score 1–10. Below 7 = NEEDS_WORK, 7+ = APPROVED. Only approve when no functional bugs remain and the goal is meaningfully met. List what works, what's missing, specific next steps."
4. If APPROVED → run Validator. If validator passes → done. If validator fails → feed errors back as next iteration.
5. If NEEDS_WORK → next iteration.
6. If max iterations hit → stop and report.

Subsequent Iterations

Skip Explorer. Pass to Implementer: "Previous critique: [concise critique summary]. Address these issues." Keep it short — don't forward full outputs, summarize in 2-3 lines.

Validator

After Critic approves, spawn general-purpose subagent. Validate based on detected project type:

• Python: activate env if provided (Windows: <path>\Scripts\activate, Linux/macOS: source <path>/bin/activate), run test command. Fallback: python -m py_compile <changed_files>
• JS/HTML: node --check <changed .js files> for syntax. If playwright/puppeteer available, open the page and check for console errors.
• Other: run the user-provided test command, or skip with a note in the final report.
• Return PASS or FAIL with error output.

If FAIL: loop back — Implementer gets the error output as its next critique.

Surfacing to Human

Only interrupt the loop to ask the human when:

• A decision has real consequences the user didn't pre-authorize (e.g., deleting data, changing APIs)
• The goal is genuinely ambiguous and the implementer can't make a reasonable choice

Do NOT surface for: implementation approach choices, minor trade-offs, or questions answerable from the codebase.

Stopping Conditions

• Critic returns APPROVED AND Validator returns PASS
• Max iterations reached
• Implementer signals nothing left to do

Final Report

Summarize: what was accomplished, iterations run, final critic score, validator result, any remaining issues.

Patterns

See references/patterns.md.