DREvo boosts LLM agent harnesses by 16% with self-evolution

LLM agents rely on 'harnesses'—the system that orchestrates prompts, tools, and memory—to perform complex tasks. Building these harnesses typically requires substantial expert effort. Recent research explored harness self-evolution, where the system iteratively proposes and evaluates improvements using historical trial experience. However, accumulated experience often leads to performance fluctuations rather than stable improvement. DREvo, introduced by a team of researchers, addresses two key limitations: historical experience is not dynamically reassessed for current relevance, and no explicit mechanism translates valid experience into directional search guidance. DREvo integrates three components—function-level evidence anchoring, state-dependent evidence recalibration, and role-conditioned search intent distillation—to determine which past evidence remains valid and where the harness should evolve next.

Evaluated across five diverse benchmarks under limited evolution budgets, DREvo consistently outperforms existing methods. It achieves average gains of 16.2% on domain reasoning tasks and 14.2% on agentic tasks compared to baseline self-evolution approaches. Crucially, DREvo produces much smoother evolution trajectories, meaning performance improves steadily iteration by iteration rather than fluctuating wildly. This makes it reliable for real-world deployment where compute budgets are constrained. The method requires no additional expert input during evolution, opening the door for fully automated LLM agent optimization. By distilling historical experience into precise search directions, DREvo turns past failures into stable, scalable progress—a key step toward autonomous AI systems that improve themselves.