The Problem

Today's agents have safety systems. They're not enough.

Claude Code has an AI permission classifier. Codex has sandboxing and graduated approval policies. Hermes has Tirith pre-exec scanning. These are real improvements over raw allow/deny lists.

But they all share three structural weaknesses:

• The agent can see the rules. Pattern lists, sandbox boundaries, approval triggers - all visible to the model. Visible rules become optimization targets.
• The agent evaluates itself. Claude Code’s classifier runs in the agent’s own context. Codex’s approval logic lives inside the same process. The entity taking the action is also judging whether the action is safe.
• Trust is static. The same rules apply whether it's 2 PM or 2 AM, a prod file or a throwaway scratchpad. The context of the request rarely changes the outcome.

Governor keeps agents effective by applying contextual trust to tool use: stopping unjustified behavior, allowing justified work, and escalating only when needed.

Governor

Trust for contextualized tool use.

Governor doesn't replace your agent's permission system. It wraps it with a separate evaluation layer, outside the agent's context, that decides whether each action is sufficiently justified: who is acting, what they're acting on, why the action is justified, and when it is happening.

The same tool call gets different treatment at 2 PM vs 2 AM, on a test file vs a production config, with a human present vs unattended. Same action. Different context. Different trust decision.

Governor is designed for high-stakes tool use where the same action may be acceptable in one context and unacceptable in another.

Works with Claude Code, Codex, Hermes, NemoClaw, OpenClaw, and custom agents.

Why Now

Every major AI lab is shipping autonomous agents.

Agents are getting more capable every week. Their permission systems are improving too - but the control logic is often still visible to the model.

Philosophy

Governor is not moral reasoning. It is contextual trust control.

Governor does not rely on static permissions or accumulated reputation. It makes a fresh decision for every action using the live situation around that request. Trust is contextual and continuous, without storing histories or reputations.

The goal isn't to moralize agent behavior. It's to keep agents useful: allow justified work, constrain risky actions, and involve humans only when the justification is not sufficient.

How It Works

The agent can't see the rules.

Governor evaluates actions through a trust architecture that is invisible to the agent. The agent doesn't know what signals are being weighed, how decisions are made, or where the boundaries are.

That's the point. A governance system the agent can observe becomes an optimization target. Keeping the evaluation layer out of the agent's context makes that harder.

Trust is a gradient, not a switch. Each action is scored from scratch against live contextual signals for tool use, and the agent does not see the scoring logic.