Coordinating many LLM-based agents on a single machine is messy: processes collide, state leaks, and informal handoffs break reproducibility. SwarmForge treats each agent as a disciplined role — isolated worktrees, per-role prompts, tmux windows, and a small handoff protocol — to make multi-agent engineering observable and repeatable.
What Sets It Apart
- Role-first topology: workflows are defined by a config-driven swarm (two-pack/four-pack/six-pack) so you pick how many quality gates and review roles you want, not the other way around — so what: you can trade speed for rigor by swapping branches.
- Local isolation with git worktrees: each role runs in its own worktree and gets synchronized helper scripts on PATH — so what: handoffs point at commits or short notes, avoiding long freeform payloads and making state reproducible.
- Observable, low-dependency runtime: tmux-backed windows or terminal adapters let you watch each agent in real time and recover windows without losing session state — so what: debugging and auditing multi-agent behavior becomes practical on a developer machine.
- Lightweight handoff protocol: agents use validated git_handoff or note messages plus helper scripts (swarm_handoff.sh, ready_for_next.sh, done_with_current.sh) — so what: coordination is explicit, validated, and durable on disk.
Who It's For & Trade-offs
Great fit if you want a local, auditable multi-agent development loop where roles (specifier, coder, refactorer, architect, QA, etc.) map to separate worktrees and observable terminals. It’s useful for teams experimenting with agent-driven workflows, TDD-driven agent coding, or reproducible multi-agent handoffs.
Look elsewhere if you need cloud-hosted orchestration, automatic scaling across machines, or a managed UI — SwarmForge intentionally keeps everything local and minimal (zsh, git, tmux, babashka, and at least one LLM backend required), so it’s not a hosted orchestration or MLOps platform.
Where It Fits
Use SwarmForge to prototype and run disciplined multi-agent development on a laptop or single server: pick a runnable branch for the desired workflow shape, let role prompts and constitution articles drive behavior, and rely on the handoff files for durable inter-agent contracts.
