Hey Product Hunt! 👋 We’re building Gear, an open-source framework for recursive self-improvement of AI agents.

An agent’s capabilities depend on more than its model. Its harness—the prompts, tools, skills, workflows, and context management surrounding the model—shapes how it approaches and completes tasks.

Gear makes that harness something the agent can systematically improve.

Give Gear a set of representative tasks and clear success criteria. It runs the agent, analyzes execution traces and evaluation feedback, proposes harness changes, and tests the resulting candidates.

Run tasks → diagnose failures → evolve the harness → evaluate candidates → retain improvements.

The model’s weights can stay fixed throughout this process. Gear records harness versions and their evaluation results, so you can inspect what changed and reuse the resulting harness.

We’re building toward a scikit-learn for recursive self-improvement (RSI): reusable infrastructure that makes improvement algorithms easier to implement, compare, and extend. Developers can customize the search strategy while reusing the machinery for execution, evaluation, and tracking results.

One result we’re excited about: bringing Luna to Astra-level performance in objective completion.

In our experiment, using AutomationBench’s 100 public Marketing tasks for both optimization and evaluation, GPT 5.6 Luna with Gear achieved 88.9% objective completion, compared with 84.1% for GPT 6 Astra with Codex—both at max reasoning effort.

We’d love for you to try Gear on your own tasks, contribute an improvement algorithm, or tell us where your current agent falls short.

GitHub · Getting started

Which agent workflow would you want to improve first?