Tigera: Pods and Ports Can't Describe an AI Agent: Write Your First Per-Call Policy

 PT
Online

Once an AI agent is discovered, traditional Kubernetes pod and port rules fall short. Agents dynamically choose models, MCP tools, and other agents at runtime over the same HTTPS connection, meaning network-level policies either block them completely or allow unrestricted access.

This session covers per-agent governance. We replace shared API keys and broad service accounts with verifiable agent identities and scoped, short-lived credentials. You will learn to write fine-grained policies, permitting specific models and read-only tools while denying other calls, and safely test them in observation mode before enforcing them in production.

What you will learn

  • Three calls, one connection. Agent-to-LLM, agent-to-MCP and agent-to-agent calls, and why layer 3 and 4 rules can’t tell them apart.
  • Identity before policy. Namespace and service account versus per-agent identity from your existing identity provider (Okta, Entra ID, Keycloak) or SPIFFE/SPIRE workload certificates. Scoped tokens in place of shared API keys.
  • Writing the first mandate. A Cedar policy with a principal, an action and a resource: permit one model and two read-only MCP tools for one agent, deny everything else by default.
  • Observe, then enforce. Running the policy in observation mode, reading the would-be denials, tightening, then enforcing. An allowed call and a denied call side by side in the decision log.