GEAR

GEAR is an open architecture proposed by a multi-institution research team for improving the harness that surrounds a large language model agent, rather than the model weights themselves. It frames recursive self-improvement as a two-level meta-learning process, with published code on GitHub and an accompanying paper.

The inner loop lets a rollout agent use feedback from a single task to refine its reasoning, planning, and actions. The outer loop employs a Meta Agent that scans experiences across tasks to spot recurring failures and effective strategies, then proposes reusable modifications to the harness. Candidate changes are evaluated and only promoted when they hold up, so the two loops can be optimized separately while sharing experience as the connecting layer.

The project targets researchers and engineers working on agent architectures, evolutionary search, and evaluation methodology. It is also relevant to teams that want a structured way to test whether harness changes generalize beyond the task that inspired them, instead of relying on one-off prompt tweaks.

Because GEAR is a research framework, expect to work with code, experiment configurations, and your own task suites. Start by reproducing the reported evaluation setup before introducing new mutation or promotion rules, and keep the inner and outer loops instrumented separately so you can attribute gains correctly.

In short, GEAR offers a disciplined blueprint for treating agent harness improvement as an emp