Engineering the harness
summary
Two teams—one with a 28‑repo brownfield platform and one greenfield agentic workflow—build a shared harness called an AIKit wrapper that centralizes skills, specs and agents to stop context drift; it adds guides and sensors (global instructions, skills, structured intake, confirmation gates, tests mapped to ACs) plus sub‑agents and progressive disclosure to keep context windows small; pipelines run autonomous red/green/refactor loops with least‑privilege agents and human approval only for high blast radius changes, and every PR includes a change summary checklist for reviewability.
AI-written summary · about
chapters (17)
- 0:00Engineering harnesses
- 1:09Harness and layers
- 4:06Large-scale agent challenges
- 6:35AIKit as wrapper
- 7:36Guides and sensors
- 8:41Best practices for guides
- 10:49Automation vs human oversight
- 11:52Preventing hallucinations
- 14:45Autonomous pipelines
- 19:26Context window management
- 22:19Sub-agents for load
- 24:10Progressive disclosure
- 25:34Persistent memory
- 26:37Reviewable outputs
- 28:27PR change summaries
- 29:31Harness principles
- 30:13Q&A
related talks
46mHarness Engineering: How to Build Software When Humans Steer, Agents Execute
22mAgents FC: Harness Agents Into a Winning Team
28mArchitecting AI-DLC for the organization
37mBeyond the Harness: A Journey Towards Adaptative Engineering
21mConvention Over Hallucination: Harness Engineering for AI-Powered...
27mAgentic Engineering: Working With AI, Not Just Using It
30mArchitecting Integrity: Responsible Software Evolution in the AI Era
22mAgentic AI for Engineering Leaders | Strategy & Execution Guide