The SourceryKit SDK adds verifiable guardrails to Python agents by making outbound actions observable, enforceable, and auditable. Its modular architecture isolates responsibilities to ensure easy integration and security.
SourceryKit intercepts an agent’s outbound HTTP calls, enforces endpoint policies, records interactions, and enables deterministic evaluation of claims.
Rendering diagram…
- Agent: The host Python application. It executes HTTP requests, submits claims for evaluation, and receives verifiable results with minimal changes to core logic.
- Bootstrap System: Handles one-time initialization, including database schema setup and resource registration, before the agent starts executing. See introduction.
- HTTP Interceptor: Intercepts outbound HTTP calls to enforce policies and record payloads. Untrusted requests are blocked before leaving the process. See interceptor.
- Database Tables:
- Trusted Endpoints — A registry of allowed endpoints used by the Interceptor to validate outbound requests. See trusted endpoints.
- Intercepts — An append-only table storing detailed records of every outbound HTTP request and response. The Provably backend uses these records to generate cryptographic proofs. See interceptor.
- Handoff Payload: A structured data format used to pass claims about external calls between agents or services. See handoff.
- Evaluator: Compares agent claims against authoritative records from the Provably backend, returning a deterministic verdict (PASS, CAUGHT, or ERROR). See handoff.
- Provably Backend: The external service acting as the single source of truth for verifying agent claims by generating proofs from the Intercepts table.