open-core · v0.1 in progress

Git for AI agents.

Every action is a commit.
Anything can be rolled back.

Agents are probabilistic. They will eventually do something plausible but wrong. You can't make them perfect — you make failure survivable.

the full race

85% end-to-end · step 1 of 10
scroll to run the race ↓

revert

the undo button for AI agents

how it works

Every action is a commit.

01

intercept

Every tool call becomes an immutable event.

Endpoint, payload, before-image, idempotency key, response, timestamp, causal chain. Nothing an agent does is invisible — it is a commit.

02

undo

Compensation is a forward-moving event, never an erasure.

Actions are classified into reversibility tiers. Inverse operations are generated automatically — the undo itself lands in the ledger.

03

timeline

Branch points, state diffs, one-click revert.

The audit trail is the product: every change, every cause, every undo — queryable like git history, rewindable to any step.

revert — action ledger recording
$ agent.call("send_email", { to: "finance@acme", amount: 5000 })
✓ recorded #ab12c3 · before-image saved
$ agent.call("approve_invoice", { id: 42 })
✓ recorded #cd34e5 · before-image saved
✗ read-back mismatch: recipient ≠ original
$ revert rollback --to #ab12c3
~ compensation emitted: revoke_email #ef56ab
✓ state restored — history intact
action recorded mismatch detected ~ compensation emitted state restored
✗ bad ✓ reverted

first 90 days

The plan, in public.

We build in the open — the roadmap is a board, the PRs are the classroom, and the docs are the textbook.

  • W1–2

    Validate before we build

    Interviews with teams running agents that touch money, health, and prod config. 10 real incident stories.

  • W3–6

    The SDK MVP — the moat

    Tool-call wrapper, before-image capture, undo executor (tiers 1–3), SQLite ledger + CLI. OSS v0.1.

  • W7–9

    Pilot-ready

    Audit exports, security one-pager, 3 pilots, incident studies published.

  • W10–12

    Harden + decide

    Fix pilot pain points, pricing draft, and a go/no-go decision with evidence.

we're learning in public

Ship agents
fearlessly.

The whole company is learning — PRs are the classroom, ADRs are the textbook, and no question is too basic. The repo has a learning path for every concept we build.

4 repos · 1 org · open core · apache-2.0