Agents FC: Harness Agents Into a Winning Team
summary
A concrete recipe for turning fragile multi‑agent systems into observable, governed teams: build a harness with runtime, observability and governance layers, split work into specialized agents (captain, striker, defender, reviewer), log every handoff, and add feedback checks so failures are caught before they propagate. A first black‑box version suffers non‑determinism; reworking it into Agent FC v1→v2 and wiring tools via Amazon Bedrock AgentCore plus a gateway, tool registry, Lambda tools and CloudWatch traces delivers three harness layers of control and repeatability.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:32Why agent teams resemble football teams
- 1:45Reliable agents need a reliable harness
- 2:04Dividing work and engineering context
- 4:02The agent loop and orchestration
- 5:38Football as a model for multi-agent systems
- 5:51Agents FC version one
- 8:16Why the black-box design failed
- 9:36Harness engineering
- 11:04Three layers of an agent harness
- 12:24What version one was missing
- 14:10Agents FC version two
- 14:55Mapping the system to Amazon Bedrock AgentCore
- 16:16AgentCore demo
- 18:02Multi-agent football simulation
- 20:23Recap
- 21:06Resources and closing
related talks
31mEngineering the harness
46mHarness Engineering: How to Build Software When Humans Steer, Agents Execute
40mHow To Tame Your Agents?
15mDevelop at Idea Velocity
32mWhat if the harness mattered more than the model?
28mThe Put That Sh*t On Agent: Building Multi-Agent AI Systems That Actually Talk to Each Other*
37mBeyond the Harness: A Journey Towards Adaptative Engineering
46mVibecoding your Own Multi-Agent Workstation