← Field Notes
NOV 6 · Clipped · via Hacker News Agent IdentityObservabilityPermissions

Zalando ships a consent layer connecting hundreds of agents to dozens of servers

Zalando put consent outside the agents. A human can see and revoke access in one place. That is what the permissions pattern needs to actually work at scale.

Machine summary of the source

Zalando built and open-sourced a delegation layer that sits between its AI agents and the internal services those agents call. The layer handles OAuth 2.0, an open standard for granting access without sharing passwords, so each agent gets a credential scoped to exactly what it needs. Hundreds of agents connect to dozens of internal servers through it today.

The design decision worth noting is that consent lives outside any single agent. A human operator grants access once, in one place, and can revoke it in that same place. The agent never negotiates access on its own. That separation means the access record is readable and auditable without opening the agent itself.

For product teams, this is evidence that permissions work better as a shared service than as logic baked into each agent. A designer building a multi-agent product needs a place where a human can see, in one view, what every agent is allowed to do. Zalando's approach makes that view possible. Without it, access decisions scatter across agents and become hard to review or reverse.

The summary above is generated; the note at the top is the editorial judgment. Primary source ↗