Your cluster has no idea who's calling.

Agents talk to APIs, call tools, hit databases & to your auth system they look like a service account, a shared API key, or a bearer token in a header.
Retrofitting identity after a fleet exists is painful.

The CNCF introduced kagenti to give every agent a SPIFFE/SPIRE cryptographic identity; kagent ships agents as K8s CRDs subject to RBAC. SPIRE has been CNCF graduated for years - the missing piece was the agent side.

This session reframes agent governance as an identity problem, not a policy problem.

For workload identity we'll demo per-agent SPIFFE IDs via kagenti, so every model call and tool call is attributable to a specific agent in a specific trust domain.

Authorization:
Kyverno policies on the agent identity at request time.
Auditing:
OTel GenAI spans tagged with the SPIFFE ID, so traces tell you which agent did what
Least privilege:
agentgateway gating MCP tool access by identity.

Identity-first, policy-second, audit-third.