🌐 revert-labs.github.io — the org site
Agents are probabilistic. They will do something plausible but wrong — delete the wrong row, email the wrong person, approve the wrong invoice. You can't make them perfect. You make failure survivable.
| Repo | What it is |
|---|---|
revert |
The open core (Apache-2.0). An event-sourced action ledger + undo engine for AI agents. SDK, CLI, docs, tests — zero dependencies. |
revert-enterprise |
The control plane (private). Cross-service compensation, compliance exports, timelines. The crown jewel. |
.github |
This org: community health files, team charter, branding. |
agent ──► [ tool call ] ──► SDK records: payload · before-image · idempotency key
│
▼
action ledger (immutable events)
│
undo engine (tiers 1–5, verify-by-read-back)
│
▼
the outside world — now reversible
85% reliability per step × 10 steps ≈ 20% end-to-end. Making agents 90% reliable still fails 65% of 10-step workflows. The fix isn't better agents — it's survivable failure. Revert is the undo button that makes autonomous agents safe to deploy.
- 🛠️ v0.1 — SDK MVP in progress (roadmap board)
- 🧪 Research-backed: replay-or-fork restore, compensation tiers, ACRFence defense
- 🤝 Open core, teaching-first culture
- Contribute: read
CONTRIBUTING.md— PR-driven, small commits, tests with everything - Learn: our
learning pathmaps every concept to the code - Ask: no question is too basic — the whole company is learning
Revert Labs — ship agents fearlessly. Every action is a commit; anything can be rolled back.