AgentSkeptic
AgentSkeptic is a developer tool that verifies whether an agent's claimed tool effects actually match the state of downstream stores (like SQL databases) at verification time. It solves the problem of trusting traces that may show success while stored data disagrees, by producing a deterministic Outcome Certificate. It is for developers building AI agents or automation that perform irreversible actions like shipping, billing, or closing tickets.
✨ Key features
- Read-only SQL checks against your database at verification time
- Outputs deterministic Outcome Certificate v3 on stdout
- Provides truth_check_verdict and release_critical_truth_check_verdict on stderr
- Supports optional coverage budgets for policy enforcement
- Can compare two saved Outcome Certificates for semantic diff
- Works with SQLite, HTTP witnesses, object storage, vectors, Mongo
🎯 Use cases
- Verify that a CRM upsert actually created the expected contact row
- Gate irreversible actions like shipping, billing, or closing tickets
- Compare trust posture between two runs using certificate diff
- Integrate into CI to catch when agent traces lie about state changes
📦 Installation
🧰 Requirements: Node.js >= 22.13 for local runs, or Docker; no API key required for the default OSS path.
npm install agentskeptic
🚀 Usage
npx agentskeptic check --workflow-id wf_example \
--project ./path/to/your-app \
--db ./path/to/readable.sqlite
⚠️ Good to know
The default path runs stateless contract verification locally; stateful enforcement (baselines, drift, acceptance) is a later opt-in commercial path requiring an API key and license server.
❓ FAQ
What does 'trusted' mean in the verdict?
It means the checked outcome matched expected downstream state, and only this verdict means the workflow can be relied on.
Do I need an API key to use AgentSkeptic?
No, the default agentskeptic check path needs no AGENTSKEPTIC_API_KEY and no license server; it runs stateless contract verification locally.
Can I use AgentSkeptic with Python or LangGraph?
Yes, the default verification contract is unchanged whether you use the npm CLI or the Python SDK/extras documented in docs/integrate.md.
How does AgentSkeptic handle telemetry?
Anonymous product-activation telemetry is disabled by default unless you opt in with AGENTSKEPTIC_TELEMETRY=1 or persist config; it never sends workflow payloads, database contents, credentials, prompts, traces, or verification artifacts.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.