Don’t Trust the Agent’s “Done”: Verify the System State

Description and use

By treating agent reports as claims and verifying them against authoritative sources, the system can prevent false positives and ensure that execution decisions are based on actual state rather than self‑report.

Problem

AI agents report success but that does not guarantee the expected real‑world state exists.

Outcome

By treating agent reports as claims and verifying them against authoritative sources, the system can prevent false positives and ensure that execution decisions are based on actual state rather than self‑report.

Workflow

• Agent performs task and returns a claim (e.g., service is running).
• Identify the authoritative source for the claimed property (e.g., systemd for services).
• Query the authoritative source to obtain evidence (e.g., ActiveState=active).
• Compare evidence to the claim to produce a verdict (CONFIRMED, NOT EVALUABLE, or BLOCK).
• Use the verdict to decide whether to allow execution or block it.