What Agents Do When You’re Not Looking
summary
Built a fleet of agents across 25 projects that run ~600 sessions, share 1.5 million messages and cost about 2 billion tokens. An orchestrator, implementers and verifiers use work trees; a bug repeatedly appears when agents act in private copies and CI overloads after 984 automated commits. A behavioral layer (Anna) clusters sessions, finds guards failing 7 times in 36 hours, shows 94% self-reporting and reduces tool error from ~77% to 2%. Harden with hooks, sensors and independent checks.
AI-written summary · about
chapters (16)
- 0:00Agent behavior gap
- 1:52Harness engineering
- 5:27Chatbots to autonomous agents
- 7:55Building agent fleets
- 10:01Agent communication and context
- 12:56Behavioral analytics
- 15:27Incomplete contracts
- 19:43Private copies and repos
- 21:53Verification with judges
- 23:21Agents without boundaries
- 25:50Agent behavior sensors
- 27:38Behavioral reports
- 28:23Agent action consequences
- 30:12Hardening costs
- 31:16Agents as living systems
- 32:00Summary and Q&A
related talks
24mAI Agent Swarms Are Amazing
15mDevelop at Idea Velocity
23mAgentic Elixir
22mThe Agent Control Plane: Managing Thousands of AI Agents
57mWhen AI Agent Teams Work (And When They Just Create More Work)
09mI Run a Fleet of AI Agents Across Three Machines. Here's What Broke.
15mThe Responsible Path to Hyperproductivity with AI Agents
40mHow To Tame Your Agents?