Harness Engineering: Same AI, Better Agent (I Built One)

Harness engineering is everything around the AI model: tools, memory, checks, guardrails. I built one that makes this channel's videos. Here's every part, and what broke. One team kept the same model and changed only the harness around it. Their coding agent jumped from the top 30 to the top 5 on Terminal Bench 2.0. This video walks through the five parts of an agent harness, using the three-agent system (maker, critic, analyst) that researches, scores and ships the videos on this channel. The five parts: 1. Tools: what the agent can touch. 2. Memory: what it knows when it wakes up. 3. Verification: who checks the work. 4. Guardrails: what it may never do. The part almost nobody enforces. 5. The feedback loop: how the harness itself gets better. Along the way: my critic caught my maker claiming a safety lock that did not exist, and then the critic itself turned out to be wrong. In two weeks, a follow-up video puts this video's critic score next to its real subscriber count, on screen. Chapters 0:00 Same AI, better agent 0:15 Harness engineering, explained 1:11 The harness I built 1:35 Part 1: Tools 2:17 Part 2: Memory 3:28 Part 3: Verification 4:19 The lock that didn't exist 4:47 Part 4: Guardrails 5:49 Part 5: Feedback loop 6:01 When the checker was wrong 6:48 Your harness in five parts 7:03 The two-week scorecard Sources LangChain, Improving Deep Agents with harness engineering (Feb 2026): https://www.langchain.com/blog/improv... Blockchain.news on the Terminal Bench jump: https://blockchain.news/news/langchai... Birgitta Böckeler, Thoughtworks, Harness engineering for coding agent users: https://martinfowler.com/articles/har... Claude Code docs, How Claude Code works: https://code.claude.com/docs/en/how-c... Claude Code docs, Automate actions with hooks: https://code.claude.com/docs/en/hooks... Mitchell Hashimoto, My AI Adoption Journey: https://mitchellh.com/writing/my-ai-a... Anthropic Engineering, Effective harnesses for long-running agents (Nov 2025): https://www.anthropic.com/engineering... Anthropic Engineering, Harness design for long-running application development (Mar 2026): https://www.anthropic.com/engineering... InfoQ on Anthropic's three-agent harness: https://www.infoq.com/news/2026/04/an... Marmelab, The State Of AI Harness Engineering 2026 (Sep 2026), including the 391-repo audit and the 481-file instruction study: https://marmelab.com/blog/2026/09/24/... ZenML LLMOps database, Extreme Harness Engineering: https://www.zenml.io/llmops-database/... Note: the OpenAI figures (about one million lines, five months, zero typed by hand) are quoted via Marmelab and ZenML. The original OpenAI page could not be fetched. Watch next: the Claude Code multi-agent build on this channel. #harnessengineering #aiagents #claudecode #agentharness #buildinpublic --- Daily AI tips for technologists — developers, tech leads, engineering managers, architects, and CTOs. Every video: a named tool, a specific use case, and a copy-pasteable artifact you can try in under 5 minutes. Subscribe for daily Claude Code, Cursor, and Codex workflows.