Researchers at UC Santa Barbara and Microsoft have built a system that searches the research literature for ways to improve AI agents, turns those ideas into code and tests which changes work. Called ScholarEvolve, it targets the software around the language model—its tools, memory and execution procedures—while keeping the model’s weights fixed.[1]
The authors report a rise from 49.6% to 63.6% in Qwen3.5-27B’s task-goal completion on AppWorld Challenge, a gain of 14 percentage points. A separate test using GPT-5.4-mini on the Telecom portion of τ²-Bench improved from 72.7% to 81.9% on the pass¹ measure. Both are results from the research team’s benchmark experiments, presented in a September 30 preprint.[1]
Improving the software around the model
An agent harness determines how a model uses its environment: which tools it sees, what information stays in context, how past work is stored and what sequence of actions it follows. Two agents using the same model can perform differently because one retrieves the right information or handles an execution error more effectively. This is the part ScholarEvolve changes.[1]
The framework divides the harness into five modules: tools, context management, skills, memory and workflows. It looks for distinct research approaches within each module, implements candidate changes and evaluates them. It then tests combinations, because a useful memory strategy and a useful workflow can interact in ways that separate tests miss.[1]
Selected results with model weights held fixed
| Evaluation | Initial harness | Evolved harness | Change |
|---|---|---|---|
| AppWorld Normal / Qwen3.5-27B | 69.0% completion | 81.4% completion | +12.4 percentage points |
| AppWorld Challenge / Qwen3.5-27B | 49.6% completion | 63.6% completion | +14.0 percentage points |
| τ²-Bench Telecom / GPT-5.4-mini | 72.7% pass¹ | 81.9% pass¹ | +9.2 percentage points |
Research becomes a source of testable changes
The distinctive step is broadening where candidate fixes come from. An optimiser driven only by failure logs can revisit the techniques its coding model already knows. A literature search can supply a different mechanism: another way to structure context, maintain state or organise a multi-step task. Evaluation then decides whether that idea improves this particular agent.[1]
The authors also outline a continuing cycle in which newly published work supplies further candidates. This makes the research feed part of software development rather than a separate reading exercise. The useful output is an implemented, measured change, with a record of which combination performed best.[1]
For engineering teams, the result supports a practical conclusion: upgrading the model is not the only route to better agent performance. The surrounding software is an experimental surface of its own. A deployment decision would still require tests on the team’s actual tools and tasks, especially operations that differ from the benchmark. ScholarEvolve supplies a method for finding those candidate improvements more systematically.
Sources & further reading
- Learning from Research: Toward Lifelong Agent Harness Evolution arXiv preprint · 30 Sep 2026 · Accessed 25 Oct 2026