When an AI coding agent fails on a real software project, the usual reaction is: "the model isn't smart enough yet." This paper pushes back on that assumption. The authors argue that a huge chunk of the reliability gap has nothing to do with the model — it lives in the harness, the often-invisible piece of plumbing that sits between the AI and the codebase.

Think of it this way: if a foundation model is the brain, the harness is everything else — the hands, eyes, and short-term memory. It decides what files the agent can see, what tools it can call (run tests? edit code? grep the repo?), how errors get reported back, how long conversations get compressed when they grow huge, and how the agent confirms its work actually fixed the bug. Two agents using the exact same model can perform wildly differently depending on the harness wrapped around them.

The authors propose treating harness design as its own engineering discipline rather than an afterthought. They sketch out what a good runtime substrate needs to do:

• Observation — give the agent useful views of a large project without dumping the entire codebase into context
• Action — expose tools that match how human developers actually work (edit, test, debug) rather than awkward proxies
• Feedback — return tool errors in a form the model can actually act on, not raw stack traces or truncated noise
• Verification — establish that a proposed change really fixed the problem, not just that it compiled

The key insight is the model–harness–environment system: capability isn't a property of the model alone, it's emergent from the whole stack. A weaker model with an excellent harness can outperform a stronger model with a clumsy one. This reframes a lot of recent benchmark debates — claims like "GPT-X solves 60% of SWE-bench issues" are really measurements of a specific model-plus-harness combination, not of the model itself.

Practically, this matters because the field has been pouring effort into making bigger and better models while the harness layer has been getting only ad-hoc attention. If the authors are right, there's substantial unclaimed performance just sitting in better runtime design — better context management, better tool ergonomics, smarter compaction, more honest verification loops. It also suggests that comparing agents fairly requires holding the harness constant, which today's benchmarks mostly don't.