An AI research method turns agent failures into upgrades — without retraining the model.
Most computer-use agents — multimodal models that click, type, and navigate software on a user's behalf — are trained on what they got right. Failed runs get discarded. Researchers have now published a complementary approach: feed those failures to a separate LLM, let it diagnose what went wrong, propose fixes, and generate code patches. Those patches, lightly reviewed by humans, are applied at inference time rather than baked into new training runs. Testing on the OpenCUA-72B model against the OSWorld benchmark pushed the task success rate from 42.3% to 48.9% — a 6.6 percentage-point gain with no additional training cost and only modest added compute.
The significance is less the benchmark number and more what the method sidesteps. Fine-tuning large multimodal models is expensive and slow; inference-time patches are neither. That makes failure-driven improvement a practical tool for teams that cannot afford continuous retraining cycles, and it surfaces a category of signal — systematic error patterns — that success-only pipelines structurally ignore.
Computer-use agents are already competitive: Anthropic, Google, and Microsoft have all shipped or announced versions. Getting from "works sometimes" to "works reliably" is the bottleneck, and mining failures for that signal is a straightforward idea that, somehow, took this long to test formally.