AI agent governance for every tool call.

AI agent governance is the policy layer that decides what an AI agent is allowed to do at runtime, on every tool call. Instead of trusting prompts to encode the rules, governance moves them into a versioned policy that the system enforces inline, allow, escalate, or block, with attribution for every decision.

In the Averta dashboard or through the API, not in agent prompts. A policy is a set of rules you attach to an agent's API key.

On the execution path, before any tool call runs, at the request, tool, tool-result, and output stages.

The intent and risk score of a request, which tools an agent may use, the contents of a tool result, and the model's output. A policy can allow, escalate, or block, restrict which tools are exposed, or rewrite an unsafe output.

A blocked call never executes, an escalated call is held for review, and the decision is recorded with full context. You can also restrict an agent to a safe subset of tools rather than blocking outright.

Each agent runs the policy you attach to it, so the same request gets the same decision every time, regardless of which agent or framework makes it.

Yes. The policy layer sits at the tool-execution boundary, independent of model and framework. Switching models or upgrading agents does not change the policy surface.

Classification scores intent and risk, the MCP Gateway exposes approved tools, and the Tool Policies Framework decides allow, escalate, or block on every call. Together they govern the full path from input to action.