agit
Git for AI agent state. Commit, branch, diff, and bisect an agent's memory and actions to replay and debug how it went wrong.
What it is
agit is version control for agent state. Instead of tracking code diffs, it commits the structured state an agent carries (memory and world state) as content addressed, hash chained commits you can branch, merge, diff, and revert.
The problem it solves
Software has git, so when something breaks you have history, blame, diff, and bisect. Agents have none of that. When an agent diverges from expected behavior there is no commit history and no way to trace which state change caused which action.
How it works
An execution engine wraps each agent action with a commit before and after, recording message, action type, author, and timestamp. The Rust core handles three way merges, Merkle tree diffing, bisect to find the first bad commit, and a causal graph that traces state dependencies across commits. Storage runs on SQLite, PostgreSQL, or S3, and prebuilt hooks integrate with the Claude Agent SDK, LangGraph, OpenAI Agents, the Vercel AI SDK, and CrewAI.
Where it fits
agit is the audit and debugging layer under an agent framework. Switchboard and Sardis use the same hash chained primitive for coordination and payments; agit applies it to the agent’s own state so a run can be replayed and diagnosed.
Status
The core version control engine (init, commit, branch, merge, revert, log, diff, bisect, causal graph) is working and tested, with a FastAPI server and a roughly twenty five command CLI. Swarm orchestration and self healing modules exist but are early.