Been building with LangChain agents and kept running into the hallucination problem — agents acting on wrong information with no circuit breaker.
Built a verification layer that sits in the pipeline and catches it before the agent acts. Wrote up a full tutorial with working code:
Short version: pip install langchain-agentoracle then add AgentOracleEvaluateTool to your agent. Returns per-claim verdicts (ACT/VERIFY/REJECT) with confidence scores and evidence.
Free to test with no setup: agentoracle.co
Happy to answer questions about how it works.