δΈδΊΊ = δΈζ―εΌεε’ι — One person, a full AI development team
Overview
Agent Swarm is a framework for running an autonomous AI development team. It connects 4 modules β daily issue scanning, task tracking, AI code review, and intelligent failure recovery β into a self-sustaining loop. You set the direction; the agents handle the rest.
Modules
Four purpose-built modules that snap together into a fully autonomous development pipeline.
Runs daily at 09:00 EST via cron. Scans GitHub Issues for new actionable items, filters already-handled tasks, and automatically dispatches them to the coding agent β all before you've had your first coffee.
Every spawned agent writes to active-tasks.json on creation and
updates it on completion. One command gives you real-time status across every
parallel task, PR URL, and outcome note.
Every PR triggers a self-hosted GitHub Actions runner on the Mac mini. It pipes the diff through Gemini CLI for a full code review covering bugs, performance, security, and quality β then posts the verdict as a PR comment and fires an iMessage notification.
Intelligent failure recovery that never blindly retries. When an agent fails, it classifies the failure into one of 4 types β missing context, direction drift, environment issue, or task too large β then rewrites the prompt accordingly before spawning again.
Architecture
A closed-loop orchestration pipeline. The main agent dispatches work; everything else runs autonomously β right up to the iMessage notification in your pocket.
You β Main Agent
Set a direction or let Morning Scan surface tasks automatically from GitHub Issues.
Main Agent β Coding Agent
The orchestrator spawns a dedicated coding agent (tmux + Claude Code) and registers the task in active-tasks.json.
Coding Agent β GitHub PR
The coding agent implements the feature or fix and opens a pull request β no human intervention.
PR β Gemini AI Review
GitHub Actions triggers the self-hosted runner on the Mac mini. Gemini reviews the diff and posts a verdict comment.
Review β iMessage Notification
Once the review completes, a notification fires to your phone. APPROVE, REQUEST_CHANGES, or COMMENT β you're always in the loop.
Failure β Ralph Loop V2
If any step fails, the failure is classified (A/B/C/D), the prompt is rewritten, and the agent is re-spawned intelligently.
Ralph Loop V2 β Failure Types
Missing Context
File not found, incomplete output, "can't see the full picture"
Direction Drift
Agent did X but you wanted Y, ignored constraints
Environment Issue
Command errors, missing deps, permission failures
Task Too Large
Timeout, chaotic output, stuck midway β needs splitting
Coverage
Every step of the development loop β from triage to shipping β handled by agents so you can stay in the architect's seat.
Morning Scan scans GitHub Issues daily and auto-routes actionable items to the coding agent.
A dedicated coding agent (tmux + Claude Code) implements every task autonomously end-to-end.
Gemini AI reviews every pull request via a self-hosted GitHub Actions runner on the Mac mini.
Ralph Loop V2 classifies failures into 4 types and rewrites the prompt before retrying.
active-tasks.json registers every spawned agent β status, PR URL, and notes in one place.
iMessage alerts fire when PR review completes, so nothing slips through without your awareness.
Coming Soon
Be the first to run your own autonomous AI development team.