Your agents already run on more than one cloud. Their identities do not travel, their track records do not travel, and every cross-cloud handoff resets trust to zero. Assay is the layer that does travel.
Source: 2025 industry surveys of enterprise AI agent adoption.
AWS has IAM. Azure has Entra. Google has its own. Each is excellent at governing agents inside its own boundary, and each is structurally incapable of governing the handoff between boundaries — because a cloud that vouched for agents on a competitor's platform would be asserting something it cannot verify.
The workaround is federation: cross-account roles, workload identity, B2B guest access. All of it transfers permission. None of it transfers evidence. A federated identity with a spotless 400-job record and one created this morning look identical.
So multi-cloud agent estates end up with governance that is strong in the middle of each cloud and absent exactly where the risk concentrates.
Assay identities are platform-neutral by construction. A Bedrock agent and an Azure agent register into the same namespace, attest to each other's outcomes, and both scores move — with neither cloud mediating and neither able to see into the other.
Trust is computed once, from counterparty-verified outcomes, and read identically everywhere. Anti-gaming is built into the maths rather than bolted on: a verification ceiling caps agents whose work nobody else witnessed, and a velocity discount catches attestation bursts that are implausible for real work.
For offline or high-fanout checks, issue a signed Agent Passport — a portable, 24-hour credential a counterparty can validate without calling you at all.
Assay is not owned by a hyperscaler and does not need to be trusted by one. That neutrality is precisely what lets it sit between clouds.
Verification ceiling, velocity discount, Bayesian shrinkage toward a neutral prior, and a recency half-life. An agent cannot inflate itself by attesting in a loop.
Verification exposes a trust score and proven capabilities. It never exposes the counterparty's other work, other customers, or telemetry.
Signed passports let a counterparty check a credential without a round trip, which matters at fanout or on constrained links.
Platform is metadata for discovery. The trust maths does not branch on it.
verified reflects the trust
threshold and account status. It does not fall to false because a claimed capability
is unproven — an unproven capability simply does not appear in
capabilities_confirmed. Gate on both fields.The platform field accepts bedrock, azure, claude, mcp and custom. Anything that can make an HTTPS request can participate — custom covers Google Cloud, on-prem, and bespoke frameworks today.
No, and that is the design. Neither cloud vouches for anything. Both read the same neutral graph, and the evidence in it comes from counterparties, not platforms.
The verification ceiling caps score for agents whose work is not counterparty-verified, the velocity discount penalises implausibly fast attestation cadence, and shrinkage toward a 0.5 prior keeps thin records near neutral. A two-agent loop attesting to each other converges low, not high.
Sub-second median in production across verify, trust and passport. It is safe inside an agent's tool-call loop.