Hi community
I see companies everywhere putting AI agents on every channel they have, and honestly it scares me a little. The agent can lie or fail while saying everything is fine, and checking it can be a huge effort. I think there should be an extra step that stops it based on evidence, before the chaos spreads.
That’s why I built Deed. The agent writes what it says it did, the real system writes the other side (the payment webhook, the database change, the customer’s confirmation), and nothing is a fact until both match. In doubt until proven. If something doesn’t match, everything that depended on it gets blocked, so the error doesn’t spread. In a simulated test with 300k actions it gave 0 false confirmations.
I’m Kerly, from Ecuador. I had never built something like this, and there were nights I felt it was too big for me. I built it for the Meta AI hackathon, for a small store that takes orders over chat.
What I’d love feedback on:
Code: [GitHub - HackMet/deed-ledger: Agent claims become facts only after the effect is read back from the real system — idempotent, fail-closed, and auditable. A temporal fact memory gates downstream decisions, so cancellations stop what depended on them. · GitHub](http://github.com/HackMet/deed-ledger)
[deed – Meta Global AI Developer Hackathon – Cerebral Valley](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4521)