AgentScore

Evaluated AgentScore against Persona and Plaid Identity before our agent-commerce launch. Went with AgentScore because the primitives match our data model — we're passing agent identities around, not person identities, and the vendors that retrofit "agent" as a subclass of "person" create weird failure modes downstream. The TypeScript SDK is well-typed, error codes are readable (agent_reputation_below_threshold beats a generic 429), and the delegated-identity pattern for sub-agents is the cleanest I've seen. Would like better docs on the reputation model — how deltas compound, what triggers a manual review. Team is responsive on Slack so it's not blocking.

Built the compliance gate on our checkout agent. /v1/verify returns in ~200ms, JWT format is standard, and the reputation delta gave us a clean way to throttle sketchy sub-agents without hard-blocking. One thing that took me an hour to sort out: the sandbox webhook signing key rotates on a different schedule than prod and the docs only mention it in the dashboard, not the webhooks reference. Otherwise solid.

Wired identity gating into our A2A payment flow. The delegated-agent identity model actually maps to how we structured our agents — didn't have to shim anything into a "person" field. Docs are terse but complete.