Identity, secrets and supply chain for agents
Agents are workloads with identities: give them short-lived, scoped credentials and treat models, tools and MCP servers as dependencies.
- Workload identity, not shared API keys: managed identities on Azure, workload identity federation on Kubernetes.
- Secrets in a vault, injected at runtime, never in prompts, repos or logs.
- Delegated access: when an agent acts for a user, use the user's token with narrow scopes (OAuth on-behalf-of flows) so audit trails name the person.
- Supply chain: pin model versions, packages and container images by digest; review third-party MCP servers like any dependency; run them with least privilege and network egress limits.
- Sandbox code execution: no ambient credentials, no network by default, resource limits.
Your Kubernetes experience applies directly: pod security standards, network policies, workload identity and image signing are exactly how an agent platform is hardened.
Practice
Written for this site to drill the lesson. Not reported interview questions.
- Design the identity model for an agent that opens pull requests on behalf of engineers.
- How would you vet and deploy a third-party MCP server that wraps your CRM?
Check yourself
Why use the user's delegated token rather than a service account?
Permissions match the user, and audit logs name the person on whose behalf the agent acted.
Saved in this browser.