The cloud native community spent five years building runtime guardrails after painful lessons. Software Bills of Materials, image signing, admission controllers, network policies, Open Policy Agent. That work assumed deterministic workloads. AI agents break the assumption.
Today's agentic pipelines fan out across Model Context Protocol tool chains, spawn sub-agents, and move data laterally through infrastructure that was never part of the original architecture review. The agent calls an API. That API calls a model. That model's output triggers another agent. Nobody authorized the full chain, and nobody can reconstruct it after the fact. I've been calling this the "openclaw moment," autonomous agent-to-agent fanout where the blast radius is invisible until something goes wrong.
This talk maps governance patterns from cloud native infrastructure onto AI agent workflows. Not as metaphor, but as implementation. I'll walk through what it actually looks like to run a multi-agent governance pipeline inside containers: deploying open-source large language models (8B to 32B parameter) on Docker and AWS Fargate with zero egress to third-party model APIs, enforcing an inference perimeter the way you'd enforce a network policy, building decision traceability as a first-class observability signal (not just metrics and traces, but what model, what prompt, what human approved it), and applying the sidecar pattern to embed governance into existing development workflows without adding a new control plane.
Concrete takeaways: a reference architecture for containerized AI governance pipelines, the three failure modes we hit that don't exist in traditional microservice observability, and an open-source inference layer (LLMesh) you can deploy today to keep model calls inside your perimeter. Whether you're a platform engineer adding AI workloads to your fleet or an AI builder who's never thought about runtime security, this is the infrastructure layer that's missing from the agentic stack.