Uncle Bob Built an AI Agent Assembly Line in Tmux

What it does

SwarmForge is a local, tmux-based orchestration layer that runs several AI agents on the same machine, each in an isolated git worktree with its own terminal session. A config file defines the swarm topology—how many agents, what roles they play, and which backend (claude, codex, copilot, or grok) each one uses. The agents communicate through a small handoff daemon (handoffd.bb) that passes validated draft files between inboxes, so one agent can finish a slice of work and wake the next without ever touching another agent’s checkout.

The interesting bit

The project treats agent management like factory-floor engineering. It ships predefined assembly-line workflows—two-pack, four-pack, and six-pack—that assign specific duties such as specifier, coder, cleaner, architect, and QA, then chain them into a strict handoff sequence. Each workflow is governed by a layered “constitution” of prompt articles, and the main branch acts as a shared factory spec while runnable branches carry the local project configurations.

Key highlights

• Config-driven topology via swarmforge.conf; no hardcoded orchestration logic.
• Per-role backend selection, so claude can specify while codex codes in parallel.
• Git worktrees under .worktrees/ isolate each agent’s state without multiple clones.
• A handoff daemon manages tmux socket access and inbox delivery so agents never message each other directly.
• Predefined workflow packs from lightweight (two-pack: coder + cleaner) to full pipeline (six-pack: specifier → coder → cleaner → architect → hardender → QA).

Verdict

If you want to experiment with multi-agent coding on your own machine, SwarmForge offers a structured, factory-floor approach that treats AI agents like a disciplined dev team rather than a chatty committee. If you just need one LLM to autocomplete a function, the overhead of tmux sessions and constitutional articles is overkill.