Agentguard puts a checkpoint in the tool call

Field notes

Agentguard inserts a policy-and-evidence check between an agent’s proposed tool call and the backend action. Its visible specimen uses an appointment update where three matching “John” records and missing caller identity produce a CLARIFY decision, no backend write, and a prescribed safe retry. The page describes a four-step rollout—replay, author, shadow, enforce—and shows an illustrative receipt-bearing endpoint change. Model and orchestration compatibility is listed broadly, including OpenAI, Anthropic, Gemini, LangChain, LangGraph, CrewAI, LlamaIndex, and custom loops.