Autonomous agents that act on a user’s behalf introduce identity and authorization risks the cloud native stack cannot adequately mitigate: data exfiltration from overprivileged or compromised agents, lost delegation context as agents cross domains and trust boundaries, and intent drift beyond authorized scope.

This talk presents a reference architecture to mitigate these risks. We use OpenClaw, an open source personal agent scoped with four added identity primitives: AGNTCY Identity Service-issued crypto-attested badges, ID-JAG tokens, declarative policy for fine-grained authorization at trust boundaries, and an end-to-end delegation chain for governance and auditing, in a Kubernetes-native stack.

Case Study: A SecDevOps engineer delegates OpenClaw to run a Kyverno config audit against a KinD workload. OpenClaw detects violations, spawns a scoped sub-agent, exchanges tokens across Keycloak realms, and posts findings to Slack after Cedar/OPA approval through Envoy AI Gateway.