git.agentic

Git for agent behavior. Atomic, reversible snapshots of the full system state that determines how an AI agent acts.

git revert knows about code. It doesn't know about prompts, tools, memory state, schema versions, or model versions — and modern AI agents are determined by all of those. When an agent regresses in production, git revert can't put the system back together.

git.agentic versions the whole tuple:

AgentVersion = (code_sha, prompts, tools, model, memory_snapshot, schema_version)

A commit captures all six dimensions atomically. A rollback restores all six coherently — including reverse schema migrations and memory state. It is the primitive that makes stateful agent systems operable.

Status

Open-source MVP complete; repo went public 2026-05-22. The 12-week build landed in full; production validation is in progress — design-partner use in real deployments is the bar before v1.0 language. The "under 5 minutes from git clone" demo claim is CI-verified: a demo CI job runs the whole broken-prompt scenario end-to-end on a fresh runner on every pull request (under 2 minutes with cached cargo dependencies). See docs/product/roadmap.md for the build's historical record.

Why this exists

Read docs/product/mvp-spec.md for the problem framing. The 30-second version:

A developer ships a "small" prompt tweak plus a memory schema change. The agent starts hallucinating. The on-call engineer reaches for git revert. It doesn't work — the schema is still bumped, the memory has accumulated contaminated rows, the tool versions have drifted. There is no git revert for the system that determines the agent's behavior.

We build it.

What ships in the MVP

• agentic CLI — init, commit, log, diff, rollback, status (plus ping, resolve, cat-object).
• agenticd daemon — Rust binary that owns the content-addressed object store and the snapshot/rollback engine.
• Python SDK (agentic-sdk) — typed client, plus a drop-in LangGraph checkpointer.
• One memory backend: Postgres + pgvector, deeply integrated.
• One framework integration: LangGraph.
• One demo: "the broken prompt".

Explicitly not in MVP scope: web UI, hosted SaaS, eval/CI/AE pipelines, MCP registry hosting, sandbox execution, A2A routing, more than one memory backend, more than one framework. See docs/adr/0001-architecture-foundations.md §9–§10 for why.

The demo

# baseline: agent works
./scripts/ask "I'm thinking about cancelling."
> "I understand. Could you tell me a bit more..."

# developer ships a "small" prompt + schema change
./scripts/deploy-bad-change.sh

./scripts/ask "I'm thinking about cancelling."
> "Absolutely! I'll cancel and refund the full amount. Done!"   # hallucinated

# git can't fix this — the schema is bumped, memory is contaminated
git revert HEAD && ./scripts/redeploy.sh
./scripts/ask "I'm thinking about cancelling."
> "Looking at your account, I see your refund processed yesterday..."   # still wrong

# the agentic way
agentic rollback v0.7
# ✓ Schema reverted          in 0.4s
# ✓ Memory restored          in 2.1s
# ✓ Prompts restored         in 0.0s
# ✓ HEAD now at i7j8k9l (rollback of v0.8 → v0.7)

./scripts/ask "I'm thinking about cancelling."
> "I understand. Could you tell me a bit more..."   # baseline restored

Full walkthrough in docs/product/demo-scenario.md.

Repository layout

git.agentic/
├── Cargo.toml                    Rust workspace
├── rust-toolchain.toml
├── crates/
│   ├── agentic-core/             content-addressed object store, snapshot model
│   ├── agentic-memory/           memory adapters (postgres + pgvector first)
│   ├── agentic-proto/            wire types for daemon ↔ SDK ↔ CLI
│   ├── agentic-cli/              the `agentic` binary
│   └── agenticd/                 the daemon
├── sdk/
│   └── python/                   `agentic-sdk` package + LangGraph integration
├── examples/
│   └── langgraph-rollback/       the "broken prompt" demo (run-demo.sh)
├── docs/
│   ├── product/
│   │   ├── mvp-spec.md           what we ship and for whom
│   │   ├── roadmap.md            12-week build narrative; repo went public 2026-05-22
│   │   └── demo-scenario.md      the canonical demo
│   ├── adr/
│   │   └── 0001 … 0017 — accepted architectural decision records
│   └── architecture/
│       ├── overview.md           system diagram and component boundaries
│       └── snapshot-model.md     the technical heart
├── CONTRIBUTING.md
└── LICENSE                       Apache 2.0

Building locally

Requirements: Rust 1.95+ (pinned via rust-toolchain.toml), Python 3.10+, Postgres 15+ with pgvector.

# Rust workspace
cargo check
cargo test
cargo run -p agentic-cli -- --help

# Python SDK
cd sdk/python
pip install -e ".[langgraph,dev]"
pytest

The broken-prompt walkthrough lives at examples/langgraph-rollback/. Everything below is rooted there:

• scripts/run-demo.sh starts Postgres+pgvector via the local docker-compose.yml, cargo builds agenticd + agentic from the workspace, then drives the commit → break → rollback cycle end-to-end.
• The Python LangGraph agent runs locally via python agent.py, invoked by scripts/ask.sh. run-demo.sh bootstraps its own Python venv (with the SDK, LangGraph, and psycopg) at examples/langgraph-rollback/.venv on first run — no manual Python setup is needed; the demo's own README covers running the agent by hand.

agenticd and the agent are both local processes, not containers.

Design partners

If you're running a stateful LangGraph agent in production and have been burned by a prompt or schema change that you couldn't cleanly revert, we want to talk. The MVP is being co-designed with three teams. Reach out to Toni (toni@git-agentic.com).

License

Apache 2.0. See LICENSE.

Further reading

• docs/product/mvp-spec.md — what we're building and why now
• docs/adr/0001-architecture-foundations.md — the structural decisions
• docs/architecture/snapshot-model.md — the technical model
• docs/architecture/overview.md — the system view
• docs/product/roadmap.md — the 12-week plan
• docs/product/demo-scenario.md — the demo