A harness is a list of things the model cannot do yet
A harness is not plumbing around a model. Every component in it encodes one assumption: a thing the model cannot be trusted to do alone. Read it that way and the harness becomes a checklist.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 11 June 2026 Updated: 2 min read
When a team asks me what a harness actually is, the textbook answer is “the orchestration around the model.” True, and useless. It tells you nothing about what to build.
Here is the version that helps. A harness is a list of things the model cannot do reliably on its own yet. Every component in it is there because someone hit one of those limits and built a workaround. Read the harness that way and it stops being plumbing and becomes a checklist of your team’s real, current distrust of the model.
The assumption behind each part
Walk a typical harness and you can name the assumption behind each part.
The system prompt and context files exist because the model does not know your codebase, your conventions or your standards. The assumption: it will not infer them correctly, so you state them.
The tool definitions exist because the model cannot reach your real systems by thinking. The assumption: it needs a hard, defined interface to act, not a vague capability.
The verification step (tests, static analysis, a build) exists because the model cannot tell its correct output from its confident wrong output. The assumption: trust nothing it produces until something deterministic checks it.
The retry and error handling exist because the model fails partway through and does not always notice. The assumption: it will not recover on its own, so the loop around it must.
The sandbox and the permission scope exist because the model has no native sense of blast radius. The assumption: it will not stop itself, so the environment has to.
A harness you can audit
Every harness component maps to one sentence of distrust. That makes the harness auditable. Go through yours and ask, for each part, what it assumes the model cannot do. If you cannot answer, that component is either cargo-culted or load-bearing for a reason nobody remembers. Either way, you just found work.
Knowing when to delete
The same reading tells you when to delete things. When a model gets better at one of these, the matching component can shrink or go. A harness that never shrinks is a harness nobody is reading.
A good harness is honest about exactly where it does not trust the model, no more and no less.
Pick one component of your harness. What does it assume the model still cannot do, and is that assumption still true?
Further reading
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Coding agent setup
- Skills hold what agents need to know, sub-agents do the thinking The first architecture decision in an agent setup is sorting what needs to think from what just needs to know. Skills carry static knowledge, sub-agents do dynamic work, and the split decides the cost.
- Agent teams win by moving complexity out of the agent Vercel, Manus and Pi got better results by removing tools from the agent. The orchestration belongs in deterministic factory steps around a small, focused agent.
- Claude Code can now navigate your code like an IDE Claude Code used to search a codebase with grep and text matching. With Language Server Protocol plugins it gets real code intelligence: references, signatures and types.
Test the constraint in your system
The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.
Training on this topic: Agentic Engineering for Developers and QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check