A behavioural harness raised a cost-optimised coding model’s success rate from three runs in ten to ten out of ten when building an API service, according to Martin Fowler. It also increased token use and runtime, but Fowler says the harnessed model still cost less on average than a frontier model without one.
Watch Desk analysis
What happened
Fowler’s comparison looked at coding agents building an API service with and without a behavioural harness. The frontier model succeeded in every run either way; the harness made the clearest difference for the cheaper model, which went from three successful runs out of ten to a clean ten for ten.
The trade-off is that using the harness took more tokens and time. Fowler argues that it works as an executable aid to attention, catching requirements drift as the agent builds. The figures come from this comparison, not a guarantee that every model or coding task will behave the same way.
Why it matters
The result offers a practical alternative to the reflex of reaching for a bigger, more expensive model whenever an agent is unreliable. A harness may help a cheaper model stick to the requirements, and Fowler’s comparison suggests the extra runtime and tokens can still leave it less costly overall.
That makes the useful question less “Which model is smartest?” and more “What setup gets this job done reliably at a sensible cost?” For teams building coding agents, the answer may involve better checks around the model, not just a more expensive model inside the box.
Our read
This is a tidy argument for measuring the whole system, rather than judging a model in isolation. The improvement is striking, but ten successful runs are a promising comparison, not a universal reliability certificate. Try the harness against your own tasks and count the added time and tokens before declaring victory.
What to watch
- Whether similar gains appear across other coding tasks and models.
- How much the harness adds to runtime and token costs in routine use.
- Whether the checks catch requirement drift without creating new failure points.
Discussion spark: If a harness makes a cheaper coding model reliable enough for the job, should teams spend first on better checks or a more capable model?
Sources and evidence
- Are Harnesses Worth the Cost? (13 October 2026, 13:35 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.