New AI Training Trick Cuts Coding Costs by Making Bots Ramble Less

Coding agents (AI that writes, tests, and fixes software by poking around a project like a junior developer) are now common in real workplaces. But every time they open a file, run a test, or rethink a step, they burn tokens — the small chunks of text an AI reads and writes, which is exactly what you pay for. A long debugging session can mean a startling bill and a long wait. A team of researchers has now published a training method called HERO that targets this problem at its root.

Earlier cost-cutting tricks tried to trim the AI's memory or cap how many steps it could take. That saves money but risks throwing away the very clue needed to fix the bug — like telling a mechanic to stop asking questions and just guess. HERO instead trains the model to work efficiently from the beginning, rather than restricting it after the fact.

The researchers noticed two useful patterns. First, successful fixes often need fewer tokens than failed ones. Second, when the AI is spinning its wheels, its internal uncertainty spikes — a signal that it's going nowhere useful. HERO uses that: it rewards actually solving the task first, then encourages shorter, more decisive paths, both across an entire job and at each individual step. In effect, it teaches the AI to think before it types.

Tested on SWE-bench Verified and SWE-bench Multilingual, two widely used sets of real coding challenges, HERO balanced success rate against token use better than leading coding agents. The catch: this is a research paper, not a product you can buy. The gains are measured on standard tests, not messy real-world codebases, and retraining a large AI model isn't something most companies can do themselves. Still, the direction is clear — the same AI help, for less money and less waiting.