AgentSkeptic is a tool focused on verifying the true system state of AI agent workflows, with the core goal of solving the silent failure problem where "the trajectory shows success, but the actual stored data is inconsistent." Its core philosophy is "trust reality, not the trajectory," by re-checking the storage that the agent claims to have changed and returning a deterministic Outcome Certificate before release.
Purpose: AgentSkeptic is suitable for scenarios after an agent or automation run where the authenticity of persisted SQL rows needs to be verified. When the trajectory looks normal but the database may not have been updated correctly, this tool can catch such problems. It is not suitable for scenarios that require proof that a tool was executed, that use log search as verification, or where read-only SQL is not the correct check method.
Main capabilities: Through read-only SQL checks, AgentSkeptic combines structured tool activity (events) with a registry to derive the expected state and compares it with the actually observed database state. It outputs Outcome Certificate v3 (stdout as machine-readable JSON), and provides truth_check_verdict and release_critical_truth_check_verdict on stderr. The tool supports multiple databases (such as SQLite, Postgres, as well as witness methods such as HTTP, object storage, vector, Mongo, etc.), and can generate certificate summaries, downloadable artifacts, and structured output in CI. In addition, it also provides a Quick Verify mode (no registry file required) and a contract mode (defining verification mappings for each tool ID through a tools.json registry).
Usage: The default path requires no API key or authorization and can run offline. Users need to install Node.js >= 22.13 or use Docker, build via npm install and npm run build, then run the agentskeptic check command and pass in the event file and registry file (default paths are ./path/to/your-app/agentskeptic/tools.json and events.ndjson). The tool supports invocation via npm package or source code, and can also be integrated in CI via a GitHub Actions composite action. For Python/LangGraph/CrewAI and other frameworks, the Python SDK or extras can be used. Advanced users can enable coverage budget policies, stateful execution (requires a commercial license), or export activation.
Applicable scenarios: Suitable for teams that need to ensure the database state is correct after agent operations, especially in critical business areas involving revenue, billing, ticket closure, etc. It can help organizations establish repeatable, auditable verification contracts, replace scattered SQL check scripts, and support locking down verification rules in CI.
Limitations: AgentSkeptic does not prove the deep causal relationships of tools, nor is it suitable for arbitrary logs or unstructured observation text. It only accepts structured tool activity (JSON or NDJSON). Quick Verify mode is temporary and is not an audited final result. Commercial stateful execution requires an API key and a licensing server. Regarding network access, local contract verification is offline by default, and anonymous telemetry is disabled by default unless the user explicitly opts in.
