The framework tackles limits of existing harness optimizers by adopting a literature-driven, multi-module approach rather than relying solely on execution traces.

Evaluations on AppWorld and tau-Bench Telecom show consistent gains across backbones, including zero-shot generalization to unseen apps and APIs.

Improvements include a tool interface with AST checks and on-the-fly API documentation printing, context management with query-aware budgeting to preserve relevant history, and Skills via Skill-as-Pseudocode for recurring API sequences.

Three rounds of literature updates without retraining yielded performance gains, though production deployment would require controlled validation and monitoring.

ScholarEvolve is a framework from UCSB and Microsoft that automatically extracts information from research papers to upgrade an AI agent's harness without retraining model weights.

Mid-task error recovery analysis shows evolved harnesses recover from intermediate tool errors more effectively than base harnesses.

Practical takeaways for enterprises include maintaining frozen backbone weights, using a security vetting gateway for literature/code, and recognizing that not all problems (e.g., policy guardrails, upstream retrieval) can be solved by harness evolution alone.

ScholarEvolve partitions the agent harness into five modules (tool interface, context management, skills, memories, workflows) to prevent cross-module breakage during upgrades.

A key finding is that optimizing multiple modules together yields stronger performance than optimizing a single module in isolation.

In benchmarks, ScholarEvolve improved Qwen3.5-27B from 69.0% to 81.4% task completion on AppWorld Normal, narrowing the gap to GPT-5.4 by about 74% without updating weights.

The optimization pipeline consists of failure auditing and literature retrieval, topic modeling (TopicGPT), blueprints and health probes, and module crossover with selective replacement based on validation performance.