The Agentic Trust Index

AI agents are being given real authority faster than anyone - outside the companies building them - can verify whether they deserve it, and what harms may result if they don’t.

The Agentic Trust Index is our answer: An independent methodology, developed jointly between GIE Foundation and KAMO Tune AI, for assessing what an agent can actually be shown to do.

Cover of 'The Trust Architect's Handbook, 2026 Edition', published by KAMO and GIE Foundation, featuring a blue background with white and light blue text about designing systems for trust, joint edition, methodology reference, identity, authority, policy, observability, and judgment.

The Methodology

Every action an AI agent takes should be able to answer four questions, and we treat an assessment as failed if any one of them can't be answered. We built the Index on to follow a single “instrument”:

Who or what actually did this, proven, not just claimed. Under whose authority, and what exactly were they allowed to do. Was that authority checked before the action happened, or did it just go through. And afterward, can it be shown to someone with no reason to take our word for it, a regulator, an auditor, a customer who got burned, in a form they have no basis to think was altered.

An assessment fails if any one of those four clauses breaks. We evaluate ten domains, set out in full in The Trust Architect's Handbook.

  1. Identity - who or what acted, proven cryptographically, not asserted by the workload itself.

  2. Authority - what the actor is permitted to do, on whose behalf, and within what limit.

  3. Delegation - whether authority narrows every time it passes to another hand, or whether the last link in the chain ends up holding everything.

  4. Secrets - what the actor can reach, with what credential, and for how long.

  5. Policy - whether the stated limit is enforced in code that runs before the action, not just written into a prompt.

  6. Isolation - what the blast radius looks like once something upstream fails.

  7. Provenance - where the model, the data, and anything retrieved into context actually came from.

  8. Supply chain - what the system inherited from packages, base images, and vendors nobody in the room built.

  9. Observability - whether the record of what an actor did could survive being handed to a regulator.

  10. Resilience - whether there's a chosen response when a control fails, and whether the kill switch has ever actually been pulled.

The ATI traces one real action end to end. We call this the Assessment Walk: Ten steps, one action, from identity through to revocation.

Much of what currently passes for AI assurance is a system telling an assessor what it would do under a hypothetical, rather than an assessor watching what actually happens when the system is put under one; a vendor's own documentation can describe a kill switch that has never been built, or a policy limit that exists in a slide deck but not in the code that runs before a payment clears.

The first real test of whether an agent's authority was actually bounded ends up taking place during the incident it was supposed to prevent, with the people affected finding out the limit was never enforced at the exact moment it needed to be. The Assessment Walk exists so that question gets answered before an agent is handling something that matters, not after.

These systems are being embedded into the infrastructure that runs public life: Hospital scheduling and triage support, benefits eligibility, court case management, power grid operations, election logistics, all aspects of finance, and the routine administration of government itself. An agent operating inside any one of those systems without a real, tested limit on what it can do is not only a product risk contained to the company that built it; it is a potentially unaligned and uncontained rogue actor inside a public institution, and the public has no way to see it, audit it, or vote on it.

A society that hands this much operational authority to systems nobody outside a handful of companies can verify is relocating a portion of its own governance to whoever built the model, without ever deciding to do so, and it will not find out where that decision was made until something inside one of those institutions goes wrong - potentially catastrophically.