brat vs LangGraph for multi-agent coding: a harness, not a framework
brat is a multi-agent *harness* for Claude Code, Aider, Codex, and OpenCode working in parallel on the same codebase. It runs on top of grite's append-only event log so coordination state survives crashes.
$ cargo install brat brat is a Rust-backed drop-in for
LangGraph (multi-agent). Same API, faster hot paths.
Install: cargo install brat.
Drives existing engines (Claude Code, Aider, Codex, OpenCode) — no rewrite.
brat vs LangGraph (multi-agent): the facts
Drives existing engines (Claude Code, Aider, Codex, OpenCode) — no rewrite.
You define the agents in LangGraph; existing CLIs are not the unit.
Append-only event log; coordination state survives crashes.
LangGraph checkpoints; recoverable, but the agent loop is on you.
Mayor + Convoy orchestration out of the box.
You build the orchestrator graph yourself.
Web dashboard at localhost:5173 for live state.
LangGraph Studio is a separate product.
Merge management for parallel agents.
You handle merges yourself.
MIT licensed.
MIT licensed.
When to use brat
- You want to run multiple Claude Code / Aider / Codex agents in parallel on one repo without them stomping each other.
- You want a crash-safe coordination state for long-running multi-agent runs.
When NOT to use brat
- You want a full agent framework where you define the graph yourself — use LangGraph.
Frequently asked questions
Does brat replace LangGraph?
No — brat is a *harness* for engines like Claude Code, not a framework for building agents. If you write a custom graph in LangGraph, brat is not for you. If you want to run multiple CLI coding agents in parallel, brat is.
How does brat handle merge conflicts?
brat ships merge-management policies (configurable). The default is to surface conflicts for human review; the strict mode is fail-fast.