Install
git clone https://github.com/anshulixyz/multi-agent-loop-kit.git ~/multi-agent-loop-kit && bash ~/multi-agent-loop-kit/scripts/install-into-repo.sh /path/to/your-repo
$
An agent-native setup kit (multi-agent-loop-kit) that shifts coding-agent work from one-off prompts to a repeating loop. A non-coding 'Beacon' agent scans repo state and ranks upcoming work into a LOOP_RADAR, generating small approval-gated task briefs. A human Project Lead approves each brief, then repo-specific agents that each own a path build the change in their own git worktree, run tests, open a PR, and journal the result before the loop repeats. It is explicitly human-supervised (not unattended autonomy) and rides on existing tools like Cursor, Claude Code, Codex, and Windsurf; it requires git 2.7+, bash, and a real project spec.
Loop
- Goal
- Every Beacon-ranked, human-approved task brief is implemented by its path-owning agent in an isolated git worktree and merged via PR, until the Loop Radar shows no approved work remaining.
- Exit when
- LOOP_RADAR.md lists no remaining approved or 'approval-needed' tasks, every open task brief has a merged PR, and the Project Lead approves no further work.
Check between iterations
bun run loop
$
Kickoff prompt
Start the "Multi-Agent Supervised Coding Loop" loop.
Goal: Every Beacon-ranked, human-approved task brief is implemented by its path-owning agent in an isolated git worktree and merged via PR, until the Loop Radar shows no approved work remaining.
Max iterations: 10
Between iterations run: bun run loop
Exit when: LOOP_RADAR.md lists no remaining approved or 'approval-needed' tasks, every open task brief has a merged PR, and the Project Lead approves no further work.
Step 1: Hand your coding agent prompts/setup-interview.md to populate AGENTS.md, ownership rules, and per-agent journals, then run bash scripts/bootstrap.sh to seed the Beacon journal and Loop Radar.
›