Open Agent Trust Specification
A common language for trust in AI agent ecosystems. OATS defines how platforms score behavior, issue portable credentials, and share reputation — across any provider, any framework, any cloud.
What OATS Defines
Trust Score Schema
Five dimensions — identity, risk, reliability, autonomy, calibration — with confidence values and standard scoring profiles. Deterministic, explainable, reproducible.
Event Taxonomy
24 standardized event types that feed into scoring. Task completion, tool calls, security incidents, moderation actions, confidence-outcome pairs. Map your agent activity once, score everywhere.
Portable Credentials
ES256-signed JWTs carrying trust claims. Any platform can verify an agent's trust status without an account or API key. JWKS-based key distribution.
Reputation Signals
Privacy-preserving behavioral reports. Hashed identities, aggregate scores, opt-in participation. The network gets smarter with every platform that joins.
Three Conformance Levels
Level 1: Trust Scoring. Level 2: Portable Credentials. Level 3: Reputation Network. Implement as much or as little as you need.
CC-BY-4.0 Licensed
The specification is open. Anyone can implement OATS. The standard belongs to the ecosystem, not to any single vendor.
Why an Open Standard?
Identity providers don't do trust
Knowing who an agent is doesn't tell you if it's safe. Identity providers verify credentials. OATS scores behavior. They're complementary.
Trust should be portable
An agent that builds trust on one platform shouldn't start from zero on another. OATS credentials travel with the agent — verifiable by anyone, anywhere.
Reputation needs a network
A single platform's view of an agent is limited. OATS enables cross-provider reputation sharing so bad actors can't hop between platforms undetected.
Implement OATS
VeriSwarm is the reference implementation — all three conformance levels, open source scoring engine, SDKs in Python and Node.js.