The Agent Control Plane: Managing Thousands of AI Agents
summary
Operating fleets of tens to hundreds of thousands of agents requires a separate control plane that provides seven control functions (lifecycle, policy, permissions, evaluation, coordination, observability, escalation); the talk shows concrete patterns — circuit breakers to enforce token and cost ceilings, canary and rainbow rollouts for new agent versions, scoped kill switches to shut down subsets of tenants in seconds, and fleet-wide anomaly detection. It demonstrates full-fidelity OTEL traces, offline/online eval pipelines, and a crawl–walk–run rollout that proved practical at scale.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 0:52Three acts of the talk
- 1:39Building agents versus operating them
- 2:28A taxonomy of agent failures
- 4:52Defining the agent control plane
- 5:42Lifecycle management
- 7:02Runtime policy enforcement
- 8:18Permissions and trust boundaries
- 9:09Evaluation pipelines
- 10:03Multi-agent coordination
- 10:42Reasoning and tool observability
- 11:18Human escalation
- 12:34Reference architecture
- 13:32Operating agent fleets at scale
- 13:52Canary and rainbow rollouts
- 14:44Scoped kill switches
- 15:33Fleet-wide cost and rate limits
- 15:58Anomaly detection
- 16:28Crawl, walk, run maturity model
- 18:41Key takeaways
- 21:47Closing
related talks
17mAI TechOps: Data and Infra for AI Foundation to Scale and Manage 1000s of Cohesive Agents
08magentctl: A CLI for Managing AI Agents & MCP Servers Like Infra
30mFuture of AgenticAI: Architecting the Global Control Plane (Keynote)
32mWhat Agents Do When You’re Not Looking
31mThe Autonomous Telco: AI Agents as First-Class Citizens in Cloud-Na...
24mAI Agent Swarms Are Amazing
19mAI Agents for Incidents: The Good, the Bad, and the Ugly
34mAI Agents: Transforming Autonomous Decision-Making