I Built a Tool That Verifies What AI Agents Actually Did

Description and use

Matrix verifies AI agent actions by cross‑checking agent reports with the actual target system, providing a confirmed, contradicted, or inconclusive verdict.

Problem

AI agents can claim to have performed actions that never occurred, leading to unreliable observability and trust issues.

Outcome

Matrix enables objective verification of AI agent actions, reducing false positives and improving trust in agent-driven workflows.

Workflow

• 1. Agent performs an action (e.g., sends an email).
• 2. Agent reports the action to the observability system.
• 3. Matrix receives the report and queries the target system (Gmail API).
• 4. Matrix evaluates the evidence and assigns a verdict: CONFIRMED, CONTRADICTED, or INCONCLUSIVE.
• 5. Matrix returns the verdict and evidence to the observability system for audit.
• 6. Users review the verification results to assess agent reliability.

Evidence

AI agents can report success for actions that never happened.

Source

Matrix checks the actual Gmail mailbox and returns: CONFIRMED — it happened, CONTRADICTED — it didn't, INCONCLUSIVE — there isn't enough evidence.

Source