Trust reality, not traces.
Agents can report success while your database is still wrong.
AgentSkeptic re-checks your real stores and blocks CI before missing writes reach production.
Read the docs · Run first verification locally
- Read-only by defaultVerifier checks your stores without writing to them.
- Structured verdict artifactsDeterministic Outcome Certificate JSON you can inspect, archive, and fail CI on.
Agent traces show what the agent attempted.
AgentSkeptic checks what actually changed.
Verification receipt
Example: Missing write
- Agent said
- CRM contact was updated
- Reality check
- Promised update missing
- Verdict
- NOT TRUSTED
- CI result
- Blocked before release
View Outcome Certificate JSON
{"schemaVersion":3,"workflowId":"wf_missing","runKind":"contract_sql","stateRelation":"does_not_match"}Paste events. Verify reality.
Run the bundled missing-write proof in under a minute. Then edit the NDJSON and rerun verification in the same session.
How it works
- Capture what the agent claimed
Your agent emits structured tool activity. - Define what should have changed
Map tool IDs to the database rows or stores they affect. - Verify against reality
AgentSkeptic re-reads the store and returns a deterministic verdict your CI can enforce.
Ready to verify your first workflow?
Start with the bundled missing-write proof, then wire the same pattern into your own CI.