Solution · Agent Governance

Authentication proves who. It does not bound how much.

The industry has largely solved proving which agent is acting. It has not solved bounding how much any single input — a retrieved document, a tool result, a prior message — is allowed to shape a privileged action before that action commits.

Status: beta. AGP works and is ready to evaluate. It sits in front of your existing agents rather than requiring them to be rewritten, so you can put it in a path and observe what it would have refused before you let it refuse anything.
The problem

Why identity is not enough

An agent with write access reads a document. The document contains instructions. The agent acts. Every identity check passed — the agent was exactly who it claimed to be. What was never bounded is how much that one document was allowed to influence what the agent then did.

Who is acting?

Solved. Agent identity and credentials are well understood.

What may it access?

Largely solved. Scoped credentials and permissions.

How much may one input shape an action?

Not solved. This is the gap the research addresses.

The approach

Influence budgets and causal receipts

The design treats every input-to-output influence as a signed compute edge carrying a per-input cap and contributing to an aggregate influence budget. A privileged action is refused before it commits when the budget is exceeded, and every decision produces a causal receipt that can be audited afterwards.

Input received
Influence measured
Budget checked
Commit or refuse
Start by running it in observe-only mode: it records what it would have refused without blocking anything, so you can calibrate the budgets against your real traffic before enforcement is switched on.
Why this matters now

The gap is in your architecture today

If you are deploying agents with write access to systems that matter, this gap exists whether or not you use our product. Identity is checked, permissions are scoped, and the question of how much a single retrieved document may steer a privileged action is simply not asked.

The useful first conversation is about your architecture, not our roadmap. Sometimes the answer is a design change on your side and no product at all — we will say so when that is true.

Give your agents boundaries

Start in observe-only mode against real traffic and see what it would have stopped.