AI agents can report success for actions that never happened.
The trace looks clean, but observability tools often rely on the agent's own account of what happened.
So I built Matrix — a tool that checks the real system instead of trusting the agent.
For example, if an agent says it sent an email, Matrix checks the actual Gmail mailbox and returns:
CONFIRMED — it happened
CONTRADICTED — it didn't
INCONCLUSIVE — there isn't enough evidence
The verdict comes with the evidence behind it.
Gmail only for now. Zero users. Day one.
I'm building this because AI agents need a way to prove what they actually did — not just report that they did it.
Try Matrix:
https://matrix-snowy-beta.vercel.app�
I'd love feedback from anyone building AI agents: How are you currently verifying agent actions?
Top comments (0)