When an AI agent produces a bug, fixing the code is no longer enough, because the flaw is likely to return. The teams delivering fastest with agents now prioritize the harness, meaning the environment that frames the model, before fixing the code.
What is a harness?
The harness encompasses everything surrounding the model: hooks, skills, linters, tests, and CI rules (continuous integration, meaning automatic code checks). These are deterministic mechanisms that test and validate the agent’s work before the code is shipped. According to the source, it’s not the model that makes the difference: a good harness with a modest model is better than an excellent model left to its own devices.
The engineering context, on the other hand, involves placing the right information in the right place at the right time. The harness is the subset that provides confidence in what is produced.
The bug, a mere symptom
Fixing the code without touching the harness is like repairing a leak without looking at the pipe. If an agent let an error slip through, it’s because a rule was missing, a hook didn’t exist, or the context was saturated. The same bug will return, in another form, in the next session.
In staging, the test environment preceding production, the method involves starting from the harness: why did it fail, at what point, which skill to modify? Often, the answer boils down to one sentence: the context was too cluttered, a piece of information was missing. If the bug is trivial, it can be fixed manually, provided it is logged so the harness learns from it.
Trace every error
For this approach to hold, you need to know where each error comes from: which model, which session, which skill, which agent produced a given commit. Without this traceability, you’re fixing blindly. With it, it becomes possible to replay the same task ten times and compare configurations, in terms of both speed and cost.
In a small team, a few developers and product profiles can send several thousand pull requests in a few months. A considerable portion of an engineer’s time, sometimes three-quarters, is then spent on the harness, which evolves with the application, with team arrivals, and with cost and security constraints.
Review and a new role
Some no longer review the majority of the code they push to production. This is not negligence, the source asserts, but a shift in trust, which no longer relies on the human eye and instead leans on an accumulation of safeguards. Pragmatism remains key: payment code is still reviewed, while a mobile app waiting several days for store validation calls for more caution. The harness level is therefore adjusted according to risk.
A static harness can also become a hindrance. Locked to an old version of a model, it can block a team for weeks who don’t understand why their colleagues are satisfied. A role still without a stable name is emerging, somewhere between devex, staff engineer, and “metabuilder.” Seniors will lock down the harness rather than review code, and product or design profiles will contribute to it.
Yesterday, we fixed what the machine wrote. Tomorrow, we fix what let it write poorly. One question remains: if the code is no longer reviewed, who reviews the harness?
