unclebob/swarm-forge

What it solves

SwarmForge coordinates multiple AI agents to work on a software project in a structured, pipeline-like manner. It prevents agents from stepping on each other by isolating them in separate git worktrees and tmux sessions, ensuring that work is exchanged through a formal, durable handoff protocol rather than uncontrolled simultaneous edits.

How it works

The system uses a "pack" or "forge" configuration to define a sequence of roles (e.g., coder, cleaner, architect). Each role is assigned a specific AI backend (such as Claude or Grok) and its own isolated git worktree.

Agents communicate via a shared runtime that manages:

• Durable Handoffs: Agents commit their work and use scripts to send it to the next role in the pipeline.
• Constitution: A set of shared engineering and workflow rules that all agents must follow.
• Local Dashboard: A web-based interface for the human operator to monitor agents, handle approval gates, and provide clarifications.
• Tmux Integration: Each agent runs in its own tmux session for isolation and visibility.

Who it’s for

Developers and project managers who want to automate complex software engineering workflows using a swarm of specialized AI agents while maintaining strict version control and human oversight.

Highlights

• Isolated Worktrees: Each agent operates in its own git checkout to prevent conflicts.
• Role-Based Pipelines: Pre-configured "packs" (two, four, or six-pack) define specific workflows from specification to QA.
• Durable Handoff Protocol: A structured system for sending, receiving, and merging work between agents.
• Human-in-the-Loop: A local dashboard allows operators to manage approval gates and answer agent questions.