Which models and benchmarks does it support?

The published evaluations use GPT 5.6 Luna and GPT 6 Astra on AutomationBench (Marketing and Support) and Terminal-Bench 2.1. The documentation describes connecting any task benchmark, but the page only shows results for these benchmarks.

Can I customize the search algorithm?

Partially. GEAR organizes evolutionary search into replaceable algorithm components (policy boundary, parent selection, proposal planning, task sampling, scoring rule, candidate ranking), and Staged GEPA is one concrete instantiation. In the current staged implementation, parent selection is exposed as a public policy hook, while evidence validation, evaluation stages, and final promotion remain controlled by the runtime.

What is the GEAR × Hitch × Rear stack?

GEAR manages optimization, Hitch provides versioned execution and evidence, and Rear supports inspection and annotation. The three tools are connected by versioned evidence and are designed to support a broader co-evolution research agenda.

A preprint dated October 2026 exists with a BibTeX entry on the site; the arXiv link is noted as coming soon.

GEAR — Evolutionary self-improvement for LLM agent harnesses

GEAR (General Evolutionary Architecture) is an open research framework for recursive self-improvement of LLM agent harnesses, developed by researchers affiliated with Imperial College London, Duke University, The Chinese University of Hong Kong, Brown University, and independent researchers. It structures improvement as a two-level meta-learning process: an inner loop where a rollout agent refines its reasoning, planning, and actions within a task, and an outer loop where a Meta Agent extracts cross-task patterns — recurring failures and effective strategies — and proposes reusable harness improvements that GEAR evaluates and promotes. Evolutionary search is organized into replaceable algorithm components, with Staged GEPA as a concrete instantiation; in the current staged implementation, parent selection is a public policy hook while evidence validation, evaluation stages, and promotion remain runtime-controlled. The framework is part of a connected stack with Hitch (versioned execution and evidence) and Rear (inspection and annotation). Published same-model comparisons report task pass-rate gains on AutomationBench Marketing (27.0% → 40.0%), AutomationBench Support (17.0% → 25.0%), and Terminal-Bench 2.1 (52.8% → 62.9%) using GPT 5.6 Luna at medium reasoning effort, with smaller gains against a GPT 6 Astra max baseline. GEAR is distributed as open-source code with documentation and a viewable example experiment; a preprint dated October 2026 is cited, with the arXiv link pending.

Key Features

Two-level meta-learning

Organizes recursive self-improvement as a two-level meta-learning process: an inner loop where a rollout agent refines reasoning, planning, and actions within a task, and an outer loop where a Meta Agent identifies recurring failures and effective strategies across tasks and proposes reusable improv

Modular evolutionary search

Evolutionary search is split into replaceable algorithm components (policy boundary, parent selection, proposal planning, task sampling, scoring rule, candidate ranking), with Staged GEPA as one concrete instantiation spanning diagnose, evolve, local, bridge, full-seed, and promote stages.

Editable harness with evaluation and promotion

Search operates over an editable agent harness covering prompts and policies, tools and hooks, skills and workflows, and context management; candidate harnesses are evaluated and promoted based on task evidence.

Replaceable policy hooks

Parent selection is exposed as a public policy hook, letting researchers choose which archived harness to improve next — e.g., sampling parents from a frozen archive of task-level specialists.

GEAR × Hitch × Rear research stack

Delivered as one connected research stack: GEAR handles optimization, Hitch provides versioned execution and evidence, and Rear supports inspection and annotation of experiments.

Benchmarked results

Reported same-model benchmark gains: AutomationBench Marketing task pass rate 27.0% → 40.0%, Support 17.0% → 25.0%, and Terminal-Bench 2.1 52.8% → 62.9% with GPT 5.6 Luna (medium reasoning effort), plus additional evaluations at higher reasoning effort.

Open research artifacts

Code, documentation, experiment records, and evaluation evidence are published, including a staged Marketing experiment that can be inspected directly.

How It Works

Inner loop — within-task learning

A rollout agent uses task feedback to refine its reasoning, planning, and actions within a single task.

Outer loop — across-task learning

A Meta Agent identifies common patterns across task experiences — recurring failures and effective strategies — and proposes reusable improvements to the harness for cross-task performance.

Evaluate and promote

Candidate harnesses are scored on tasks; successful changes are promoted into the archive for subsequent tasks, with the two loops connected by experience and independently optimizable.

Evolutionary search over harnesses

Evolutionary search runs through replaceable components such as parent selection (Staged GEPA samples parents from a frozen archive of task-level fronts) and improvement stages (diagnose, evolve, local, bridge, full seed, promote).

Pros & Cons

Pros

Fully open: source code, documentation, and a real staged Marketing experiment with its evaluation evidence are publicly available
Modular research design — search components like parent selection are replaceable policy hooks, making different search strategies directly comparable
Reports concrete same-model benchmark gains: AutomationBench Marketing task pass rate up 13.0 pp, Support up 8.0 pp, Terminal-Bench 2.1 up 10.1 pp (GPT 5.6 Luna, medium effort)
Part of a connected stack (GEAR for optimization, Hitch for versioned execution and evidence, Rear for inspection and annotation) supporting reproducible experiments
Makes reasoning effort explicit in all comparisons, which strengthens the credibility of the reported numbers

Cons

Who It's For

Best for

AI researchers studying recursive self-improvement, evolutionary search, and meta-learning for LLM agents
Engineers who want to systematically optimize agent harnesses (prompts, tools, skills, workflows, context management) with reproducible, ver
Teams benchmarking agents on tasks like AutomationBench or Terminal-Bench who want measurable pass-rate improvements under controlled same-m
Open-source contributors looking to swap in their own search strategies via the framework's policy-module boundaries

Not ideal for

Teams looking for a hosted, commercial product with pricing, support, or SLAs — none exists

Use Cases

Improving agent task pass rates on coding, marketing, and support benchmark tasks

Researching and comparing evolutionary search strategies through interchangeable policy modules

Building reproducible experiment records with versioned execution evidence (Hitch) and annotation (Rear)

Running the documented Marketing experiment as a template for new studies

FAQ

What is GEAR?

GEAR (General Evolutionary Architecture) is an open research framework for recursive self-improvement of LLM agent harnesses. It uses evolutionary search plus a two-level meta-learning process: an inner loop refines an agent's reasoning, planning, and actions within a task, while an outer loop Meta Agent extracts cross-task patterns and proposes reusable harness improvements that are then evaluated and promoted.

Is GEAR open source?

Yes. The source code is available at github.com/rsi-gear/gear, along with documentation for running a first experiment and a viewable staged Marketing experiment with its evaluation evidence. No pricing or hosted service is mentioned — it is a research project.

How much does GEAR improve agent performance?

Reported same-model comparisons show AutomationBench Marketing task pass rate rising from 27.0% to 40.0%, Support from 17.0% to 25.0%, and Terminal-Bench 2.1 from 52.8% to 62.9% with GPT 5.6 Luna at medium reasoning effort. Against a stronger GPT 6 Astra max (Codex) baseline, the gains are smaller: +4.0, +4.0, and +0.9 percentage points respectively.

GEAR