AI Agent Harness Engineering: Separating the LLM from the Deterministic Orchestration Layer
Mitchell Hashimoto's "Agent = Model + Harness" framework. The reason 88% of enterprise AI agents never reach prod isn't the prompt — it's the missing deterministic orchestration layer. Four CS primitives you can't skip: state machine, idempotency, DAG, DLQ.
Brain (LLM) only decides what task next + evaluates quality. Body (Harness) handles everything else deterministically: state / retry / DAG / DLQ.
[ Log in to see the full prompt ]Sign up free to see the full prompt, copy it, save it, and join the discussion. Free content unlocks on login; Pro content is a separate subscription.
The author expands Mitchell Hashimoto's "Harness Engineering" concept into an implementable SOP. For AI agent developers, this is the root answer to "why is my prototype amazing but my production dead" — don't spend 6 months tuning the prompt, spend 1 week building the harness.
AI agent 在 prod 穩定運行 — task 不會被 double-execute、信用卡不會被重刷、failed task 進 DLQ 給你看不是消失、agent 100% 受 harness 控制不自己亂跑
A description of the intended result. Check this entry's testing notes and try the prompt with your own inputs.
Suno Engineer's Mindset: 4 Steps to a Song That Doesn't Sound Like AI
A studio engineer's breakdown of Suno's fatal weaknesses (fried vocals, high-frequency artifacts), plus a 4-step DAW workflow and a Suno Studio cleanup prompt.
5 Claude Weekly Workflows That Stuck After 6 Months
Proposal generator / meeting processor / content repurposer / Friday review / shutdown reset — out of 40 I tried, only these 5 survived, each saving 30+ minutes per run.
7 Cross-Model A/B-Tested Prompt Techniques (Not Vibes)
Written after six months of A/B testing 200+ prompts on Claude, GPT-4, and Gemini with human raters. Covers persona + constraint stacking, anti-examples, role-reversal QA, cognitive scaffolds, emotional priming, uncertainty CoT, and steelman-first.
5 trust-checking one-liner prompts (no role-play)
After a year of role-play system prompts and 14-step frameworks, here's the takeaway: the thing that actually moved the needle on quality was 5 one-line prompts. No roles, no markdown, no "you are an expert."