AI agents just got a bit better at learning from their own mistakes - without anyone retraining the underlying model.
Researchers describe a system called Evo-Harness that lets a frozen large language model agent, meaning its weights never change, improve over time by editing an external playbook, or harness, instead. The core trick, called context-to-harness skill compilation, takes the messy, one-shot record of a single task attempt and distills it into a reusable skill that can be applied to future tasks, even across different domains. The team tested the approach on five realistic benchmarks covering coding, terminal use, and web navigation, including SWE-bench, TerminalBench2, CL-Bench, and WebArena-Infinity. They frame the underlying problem as online harness learning: in real-world settings, most tasks only give an agent one shot to learn from, and that one shot is full of noise tangled up with anything genuinely useful.
Most agent self-improvement research either retrains the model outright or relies on loose memory dumps that agents rarely use well in practice. By keeping the model frozen and editing a structured, external harness instead, Evo-Harness sidesteps the cost and risk of fine-tuning while still claiming measurable gains from a single pass at a new task. That distinction, learning without retraining, is the kind of engineering choice that could matter more for production agents than another benchmark score.
It is also, worth noting, a single paper grading itself against its own benchmark suite - not an independent audit, so treat the results as a promising first data point rather than a settled case.