Agentic Frameworks

For LangChain, Vercel AI SDK, Mastra, OpenAI Agents SDK, LlamaIndex.TS, and VoltAgent, governance is a wrapper: call visiq() once on the executor at construction time and the harness instruments the agent's tools and retrievers in place. For OpenClaw it is a native plugin that registers hooks on the daemon and covers built-in and MCP-loaded tools through the same seam.

There is no proxy to route through and no rewrite of the agent. The framework keeps doing what it does; the harness sits on the seams the framework already exposes.

Every tool call is observed with its full arguments before it executes, and a deny stops it there, pre-execution, with hard-stop semantics. Sensitive actions can route to a human instead of resolving automatically.

Every retrieval is intercepted between the source and the model. Policy can drop a document entirely or rewrite it to mask secrets and personal identifiers, and the masked version is what enters the model's context.

The harness evaluates in-process against one locally cached rule bundle, refreshed in the background. Measured overhead is under a millisecond per tool call and per document; a cold fetch against a fresh bundle is the exception, and even that is tens of milliseconds.

The only decision that waits on the network is the one that should: a human approval.

A blocked call does not crash the agent. The harness returns the denial as a structured tool result, so the agent reads it and reasons over it on its next turn.

If the harness itself cannot evaluate, agent-side behavior defaults to fail-open with loud telemetry, and teams that want the stricter posture opt into fail-closed. A policy deny enforces in either mode. The per-framework capability grid, with what each harness observes, blocks, and redacts, is published on the agent security matrix.