Three tiers of agent governance

Most writing on agent safety treats one agent at a time: what it can be made to do, what it should be prevented from doing. A report posted this week starts from the observation that the interesting failures are not inside an agent but between agents, and that the ability to govern them collapses at a specific boundary — the edge of the organisation.

The organising idea is a tier defined by the minimum governance that binds any two interacting agents, rather than by capability or by sector.

• Singular governance — one organisation deploys and governs every agent in the interaction. Controls exist and someone can apply them.
• Federated governance — several organisations deploy into a shared environment under rules they have agreed. Controls exist but require agreement to change.
• Open environments — agents meet counterparties with no central authority, where shared standards are adopted voluntarily if at all.

Within each tier the report works through risk factors, failure modes and available controls, and then asks a question that most safety writing skips: who is actually positioned to apply each control. Where the answer is nobody, it says so and characterises the gap as one requiring collective action rather than a better internal policy.

That framing matters for anyone deploying agents commercially. Inside the perimeter, an organisation can log, sandbox and revoke. Once its agents transact with a partner's agents, and then with unknown counterparties on the open internet, no single party can see the whole interaction, and controls written for the first case quietly stop applying. The report's contribution is not a new control but a way of telling which of the existing ones still work at a given distance from home.