When Software Becomes a User: Identity and Authorization for Agents in Production

Identity systems assume a human is behind every credential: one principal, a stable role, a bounded session, and someone accountable afterward. Agents hold none of these. They act on behalf of many users, call tools nobody enumerated in advance, spawn subagents, and run for hours without supervision.

So teams pick one of two bad options. Give the agent a long-lived service account and hope. Or scope it down until it can't finish the task it was built for.

Authentication isn't the hard part. Authorization is: what the agent may do, under whose authority, with which credentials, and for how long.

Nancy Wang builds identity security for agents at 1Password, after a decade scaling data protection and infrastructure at AWS, Rubrik, and Google. This keynote covers why the existing models break, how to think about delegated authority across a subagent chain, and what has to be settled before an agent touches production.

Key Takeaways

1. Which assumptions in human identity models agents break, and why service accounts and static keys make the blast radius worse.
2. How to scope authority that narrows as it's delegated, and what it takes to keep an audit trail through multi-hop tool calls.
3. How to give an agent enough access to finish a task without putting the underlying secret in the context window.