AI agents may not need a new model every time they fail. Google Research has open-sourced Harness², a framework designed to help agents improve the systems built around their models without retraining the underlying model or changing its weights.
The idea is simple but powerful: instead of modifying the model, Harness² rewrites its harness. That harness includes the prompts, tools, memory, skills, and scripts that guide an agent through a task. Google says the framework can edit eight parts: the system prompt, guardrails, memory, skills, tool guidance, scripts, subagents, and plugins.
This creates a different path for improving AI systems. The model stays in place, while the surrounding software learns what works and what fails.
Two recursive loops improve the agent
Harness² uses two levels of recursion. One loop focuses on improvements within a task, while the other improves the harness across tasks. The top loop runs on every task, and the improvement harness remains available between tasks instead of starting from scratch each time.
The framework learns from behavioral contrasts between executions, without ground-truth feedback. In practical terms, it compares how different versions of an agent behave and uses those differences to guide changes. That gives the system a way to improve even when it does not have a labeled answer for every task.
Harness² can adjust the parts of an agent that shape its actions, including its tool guidance, memory, scripts, and subagents. The model itself remains untouched, which separates improvements to the agent’s working setup from improvements to the model’s learned knowledge.
Google tested GLM-5 and Gemini 3.5 Flash across four model-harness configurations and three benchmarks. A single step, marked as k = 1, improved both models over their base harnesses.
- WorkBuddy-Bench: performance improved by up to 11.5 points across office, web, and code tasks.
- LAB: the partial pass rate improved by up to 11.8 percentage points on legal work from Harvey’s harvey-labs.
- JobBench: performance improved by up to 4.5 points on professional work tasks.
Those results show why the harness matters. A model can have the same underlying capabilities, yet perform better when its prompts, tools, memory, and scripts work together more effectively.
ScholarEvolve brings research papers into the harness
A related framework takes the idea in another direction. Researchers at the University of California, Santa Barbara, and Microsoft developed ScholarEvolve to extract information from research papers and translate it into code upgrades for an agent’s harness, also without retraining the model.
Current harness optimizers often depend on execution feedback and the parametric knowledge already stored in the underlying model. Meta harness evolution is reactive, so it depends on situations an agent has already encountered. ScholarEvolve addresses that limit by searching research literature for techniques that may help explain and fix failures.
The system works through four stages: failure auditing, topic modeling, blueprint creation, and module crossover. It divides the agent harness into five modules with strict input and output interfaces, allowing improvements to move between modules without changing the model’s weights.
On the AppWorld Normal test set, ScholarEvolve lifted the open-weight Qwen3.5-27B model from 69.0% to 81.4% task completion. That closed roughly 74% of the performance gap with GPT-5.4 without updating the model’s weights.
The improvements were tied to specific problems. On Qwen3.5-27B, ScholarEvolve fixed blind API calls, prevented lost observations, and stopped the agent from reinventing API sequences. On AppWorld, task goal completion rose by 12.4 percentage points, while scenario goal completion rose by 20.2 points.
Better harnesses can change model rankings
The gains continued across other tests. On Telecom, ScholarEvolve lifted GPT-5.4-mini from 72.7% to 81.9%. On the AppWorld challenge, Qwen3.5-27B jumped from 49.6% to 63.6% task goal completion and beat Stanford’s Meta-Harness by 9.0 points on unseen applications.
The results also show that combining improvements matters. Standalone skills upgrades scored 75.4%, while combined modules reached 84.8%. That difference suggests an agent benefits when separate parts of its harness support one another instead of improving in isolation.
The evolved harness also handled failure better in production environments. After an intermediate Python error, Qwen3.5-27B completed tasks 59.9% of the time with the evolved harness, compared with 44.9% for the base harness.
Harness² and ScholarEvolve point to the same shift in AI development: progress does not have to come only from larger models or new training runs. Prompts, tools, memory, skills, scripts, and research-based code changes can give an existing model a more capable way to work.
Based on
