Harness Engineering: How to Build Software When Humans Steer, Agents Execute
summary
Claims that agents do the full engineering job: repositories built and migrated by firing off dozens of agents, daily consumption measured in billions of tokens, and work done with GPT‑5.2 and Codex plus model features like GPT‑5.4 auto‑compaction. The stack uses ~5–10 core skills, bespoke ESLint rules, tests (e.g., 350‑line file limits), and reviewer agents in CI that assert retries, timeouts, and QA plans; many migrations finish in parallel and the harnesses run until tests are green.
AI-written summary · about
chapters (18)
- 0:00Harness engineering intro
- 1:09Code is free agents build
- 3:19Why agents took over
- 8:35Agent NFRs defined
- 11:26Prompt injection guardrails
- 15:20Operationalizing agent workflows
- 18:52Answering audience questions
- 20:15Live agent workflow demo
- 23:50Balancing harness complexity
- 26:39First-party harnesses like Codex
- 28:26Agent platforms collaboration
- 30:10Getting started with agents
- 32:38Agents in transit
- 34:03Scaling agent knowledge
- 36:31Automated code review
- 39:40Token budgets management
- 41:52Is code disposable?
- 43:36Future of harness engineering
related talks
31mEngineering the harness
46mWhen You Run Out of Requirements: What Happens When You Go All-In on Agents
01h 02mPlatform Engineering in the age of Generative AI
40mHow To Tame Your Agents?
15mDevelop at Idea Velocity
22mAgents FC: Harness Agents Into a Winning Team
21mConvention Over Hallucination: Harness Engineering for AI-Powered...
37mBeyond the Harness: A Journey Towards Adaptative Engineering