Gear is infrastructure that lets agents continuously learn from real-world task trajectories, adapting both their models and agent harnesses to the environments they operate in. It uses LLMs as the agents being improved, addressing the gap between general benchmark-optimized models and what specific applications need. It targets developers building agents for concrete environments rather than broad benchmark performance.
Visit project ↗Gear is an open-source RSI infrastructure for tuning agents to target scenarios. Users prepare a benchmark of representative tasks with success criteria, and a Refine Skill uses evaluation results to iteratively improve the agent's model, instructions, tools, and workflows. It targets developers wanting cheaper, better-performing agents on concrete task sets.
Gear is an open-source RSI infrastructure for tuning agents to target scenarios. Users prepare a benchmark of representative tasks with success criteria, and a Refine Skill uses evaluation results to iteratively improve the agent's model, instructions, tools, and workflows. It targets developers wanting cheaper, better-performing agents on concrete task sets.
Researchers introduce ScholarEvolve, a framework that draws on state-of-the-art research publications to automatically evolve an agent harness — the software governing tool use, memory management, and task execution — while keeping the underlying language model fixed. It organizes evolution directions into functional modules, applies topic modeling to identify improvement strategies, and evaluates their combinations. Experiments show gains on AppWorld and Tau2-Bench, and the framework ingests new publications over time for proactive lifelong evolution.
Researchers introduce ScholarEvolve, a framework that draws on state-of-the-art research publications to automatically evolve an agent harness — the software governing tool use, memory management, and task execution — while keeping the underlying language model fixed. It organizes evolution directions into functional modules, applies topic modeling to identify improvement strategies, and evaluates their combinations. Experiments show gains on AppWorld and Tau2-Bench, and the framework ingests new publications over time for proactive lifelong evolution.
A startup called Poetiq, founded by former DeepMind researchers, published a paper on harness-in-the-loop learning, where user-constructed harnesses are optimized for both immediate agent performance and the quality of execution traces used to train future models. A separate open-source project demonstrates that a harness trained on a small Qwen 3.6 35B model improves Terminal-Bench scores for four other models.
A startup called Poetiq, founded by former DeepMind researchers, published a paper on harness-in-the-loop learning, where user-constructed harnesses are optimized for both immediate agent performance and the quality of execution traces used to train future models. A separate open-source project demonstrates that a harness trained on a small Qwen 3.6 35B model improves Terminal-Bench scores for four other models.
Researchers propose HarnessCompass, a framework for automatic harness evolution that improves agent performance by iteratively refining the harness from agent-environment interactions. It addresses overfitting, reliance on trajectory-derived signals, and component interference in existing methods.
Researchers propose HarnessCompass, a framework for automatic harness evolution that improves agent performance by iteratively refining the harness from agent-environment interactions. It addresses overfitting, reliance on trajectory-derived signals, and component interference in existing methods.
A new framework called Turbo Harness adapts a globally optimized agent harness to individual task instances. It recycles artifacts from a completed global harness optimization run, summarizes them into a structured playbook, and trains a harness editor that applies instance-specific patches at inference time. This targets a limitation of prior harness optimization methods, which apply a single global harness uniformly regardless of whether it suits each instance.
A new framework called Turbo Harness adapts a globally optimized agent harness to individual task instances. It recycles artifacts from a completed global harness optimization run, summarizes them into a structured playbook, and trains a harness editor that applies instance-specific patches at inference time. This targets a limitation of prior harness optimization methods, which apply a single global harness uniformly regardless of whether it suits each instance.