Invariant

Invariant is a pre-execution control layer for production agents. It is built for teams whose agents write to real systems such as CRMs, databases, APIs, and workflows, and it is meant to stop bad tool calls before they run. The product’s stated focus is on actions that might be based on stale or conflicting state, which can lead to expensive errors.

Its core workflow is a validation gate in front of execution. Every tool call passes through five steps: the agent proposes an action, shared state is checked, contradictions such as stale facts, conflicts, or missing information are detected, a verdict of VALID, RISKY, or BLOCKED is returned, and a full reason trace is attached. The site gives an example of a CRM update being blocked after Invariant finds that a deal was already closed by a parallel agent, and it says this prevents a double-close and protects a downstream billing workflow. It also describes an HTTP endpoint, POST /actions/validate, and says the service can return exact reasons before the action runs. Invariant further says it ingests claims and tool outputs via REST or SDK, maintains a shared world state graph in real time, surfaces contradictions as first-class objects, propagates constraints, and updates a coherence score continuously.

Invariant says it works with any LLM or agent framework, including LangChain, CrewAI, and AutoGen, or through raw API calls. The page names TypeScript and Python SDKs, and says it can be self-hosted or used through Invariant Cloud. It also describes the system as low-latency validation designed for hot paths. The intended audience is teams running production agents in enterprise and operations workflows, especially agents that update CRMs, send escalations, trigger billing workflows, or modify data. The pricing page is linked on the site, and the page includes a call to start validating free.