An engineer's guide to harnesses
The state of harnesses today, how to optimize them for performance, and where they are going. A collection of essays read over 250,000+ times on X.
Harness engineering will be an essential skill in the next decade to work effectively with AI.
If the model is the source of intelligence, then the harness is what makes that intelligence useful. A harness is the scaffolding that surrounds the model to manage context in an inherently stateless LLM, handling tool calls, I/O, guardrails, and more.
Harness engineering has become one of the most important disciplines, whether you’re building a new AI application, consolidating your team’s coding setup to save on cost, or want to push any open or closed model to frontier performance.
Common questions
A harness is everything wrapped around a model that turns raw intelligence into useful work. Mechanically, it's a while (have next message) do {tool} loop.
Claude Code, Codex, OpenCode, and Cursor Agent are harnesses. So is the custom scaffolding your team wrote last quarter. We go over the fundamentals of how to get the most out of your harness, the properties behind the best ones on the market, and what we think the future looks like.
No. We think that the line of what an engineer is has become blurred, as everyone is using harnesses to some extent. The guide explicitly covers topics for non-engineers, too: GTM, ops, and knowledge workers whose job is being a good context engineer inside an existing harness. Sections on progressive disclosure, RPI, and subagents apply to anyone driving an agent. There is also deeper material on model routing, caching, and joint post-training that assumes more technical background for an engineering audience.
Engineers used to argue about IDEs; now they argue about harnesses. The model you pick matters less than most people assume, because frontier labs post-train models and harnesses together — swap either half and you pay for it in quality, latency, and cost.
The practical consequence is that your engineering judgment now shows up in the harness, not the model. One smooth harness amplifies the speed and quality of everything generated downstream.
Harness Engineering includes three main sections. Section 1 covers how to optimize the harness you already use. Section 2 is a primer on choosing one, including the buy/customize/build ladder and eight properties. Section 3 argues that harness fleets are inevitable and makes the case for depots as the next abstraction layer.
AI was only used as an editor. All research, ideas, and phrasing are our words only and reflect what we learned at Baseten while supporting frontier application-layer companies in building harnesses.
Alex Ker
Engineering + Growth
Zak Keener
Forward Deployed Engineer
