AgentGit

The system of record for a company's AI agents. Version every agent, catch silent regressions on every change, and always know whether each one still works.

early / pre-release · Apache-2.0 (open-core) · self-hostable

Working name. "AgentGit" is a placeholder — think GitHub for AI agents. It may change.

---

Why

Companies ship LLM-powered agents faster than they can manage them. A single org can already run ~100 across teams — overlapping, siloed, with no shared record of what exists, what each costs, or whether any of them still works. Most agent projects fail on the operational layer — versioning, evaluation, regression-catching, governance — not on model quality.

Two people feel two different pains:

• The Builder changes an agent's prompt, skills, tools, or logic — and has no reliable way to know they didn't make it worse. Agents regress silently; you find out from a user, or too late.
• The Consumer wants to reuse an agent another team built — but can't discover it, trust it, or tell if it's reliable, so they rebuild what already exists.

No product today owns the record of what every agent is and whether it works. AgentGit is that record — the map the rest of the tooling plugs into.

What it does

Connect the repo your agents already live in (no rewrite). From then on, on every change, AgentGit:

1. Checkpoints the agent — an immutable snapshot of everything that defines its behavior (prompt · skills · tools · code) plus its measured quality.
2. Evaluates it and attaches the score to that checkpoint.
3. Catches regressions — flags when a change lowered quality and shows exactly which cases now fail.
4. Diagnoses — explains why it likely regressed and what to change, grounded in the diff and the failing traces.

Over time, those checkpoints become a searchable map of every agent an organization runs — with health, cost, and ownership.

The core idea: the Checkpoint

Everything is built on one object. A Checkpoint is an immutable snapshot of an agent at a commit, capturing what it is (prompt, skills, tools, code) and how good it is (eval result, cost).

This is what Git structurally can't give you: a version that knows its own score.

Every feature is one of: produce a checkpoint, evaluate one, compare two, or reason about the delta between two.

How it works

  push / PR  ──►  checkpoint  ──►  run eval  ──►  score attached
                                                     │
                                        compare to last good
                                                     │
                                regression? ──►  which cases broke  +  why (diagnosis)
                                                     │
                                          surfaced on the PR & timeline

Evals run either in your CI (a lightweight GitHub Action posts results back) or in an isolated hosted sandbox — your choice, your keys.

Roadmap

Phase	Focus
1 — Builder wedge ← current target	Connect → quality timeline → catch a regression on a PR → get told why.
2 — The map	Org-wide agent registry, dependency/blast-radius graph, cost, teams & roles.
3 — Front door	Cross-team discovery, monitoring, orchestration, governance, enterprise.

Tech stack

TypeScript end-to-end across the web tier: Next.js frontend · NestJS + Fastify API · Drizzle + Postgres (row-level tenant isolation) · NATS JetStream (event bus + queue) · Octokit (GitHub App) · Anthropic (Claude) for diagnosis · a Python/FastAPI sandbox for hosted evals. All open source; the core is self-hostable by construction.

Documentation

This repo is designed to be AI-native — optimized for heavy AI-assisted development.

Where	What
docs/spec.md	The full technical specification — the source of truth.
docs/decisions.md	Register of the heavy, one-way-door decisions and their status.
docs/conventions/	How we build: data model, API & versioning, events, auth, security.
docs/architecture.md · docs/glossary.md	System shape · domain vocabulary.
changes/	Dated decision records (ADRs).
AGENTS.md	Canonical operating manual for AI agents — the read-before/write-after working loop.
sprints/ · features/	Work tracking: current scope · one folder per feature (spec + plan + worklog).

Status

Early and pre-release. Foundations and architecture are locked; implementation is starting. There is no runnable product yet.

License

Open-core: the core will be Apache-2.0 and self-hostable; team/enterprise features are commercial. See changes/2026-07-06-open-core-self-host.md.