A new technique lets AI agents stop re-thinking the same decisions every time and just run code instead.
Researchers built a system called Growing Harness that starts from a bare-bones scaffold exposing only model and tool interfaces, with no built-in strategy. When a task fails, an optimizer traces the failure to a specific chunk of code, rewrites it, and keeps the fix only if it doesn't break capabilities that already worked. Over time those fixes accumulate into a persistent harness that handles routine control decisions without calling the language model at all. Tested across two benchmarks, BrowseComp-Plus and WebArena-Verified, with models ranging from 4 billion to 120 billion parameters, it beat a standard tool-calling baseline in five of six benchmark-model settings and trailed the best result by less than a percentage point in the sixth.
The bigger story is cost and model size: Growing Harness cut the number of LLM calls by 76 to 91.8 percent and inference cost by up to 98.6 percent, while its success rate held steady at 44.7 to 45.3 percent across all three model sizes. The standard tool-calling approach collapsed to 6.7 percent success when paired with the smallest 4 billion parameter model, meaning the new harness let a much cheaper model do a much bigger model's job.
It's a good reminder that a lot of what gets marketed as 'agent intelligence' is really just the same control decisions, made over and over, that only need to be figured out once and then written down as code.