Open infrastructure for AI agents that learn from real-world experience.
GEAR is an open architecture for continuous agent evolution. We are building the infrastructure to version agent harnesses, run reproducible evaluations, turn trajectories and feedback into optimization signals, and inspect every decision from experiment to deployment.
Our current focus is harness evolution: test exact revisions under controlled conditions, preserve the evidence behind every result, and promote only the changes that earn it.
Read our recursive self-improvement (RSI) overview for the research goal, the evaluation loop, and the limits of current experiments.
| Project | Layer | What it does |
|---|---|---|
| Gear | Evolution algorithms | Generates candidate harness revisions, evaluates them under paired conditions, selects a research population, and promotes a deployment champion. |
| Hitch | Versioned execution & evaluation | Runs Codex, Claude Code, Pi, OpenCode, and DeepSeek Harness from exact versions or commits, preserving artifacts, trajectories, logs, and evaluation evidence. |
| Rear | Experiment analysis | Provides a read-only workbench for exploring Gear experiments, comparing candidates, and inspecting verified Hitch trajectories. |
flowchart LR
T["Tasks + current harness"] --> G["Gear<br/>generate · evaluate · select · promote"]
G -->|exact rollout requests| H["Hitch<br/>resolve · run · record"]
H -->|verified trajectories + evidence| G
G --> C["Research population + champion"]
G --> R["Rear<br/>inspect · compare · audit"]
H --> R
The responsibilities stay deliberately separate: Gear owns experiment state and refinement decisions, Hitch is the source of truth for execution evidence, and Rear never mutates either system.
- Want reproducible agent runs and evals? Start with Hitch.
- Want to evolve an agent harness from evaluation feedback? Start with Gear.
- Want to inspect experiments and compare their evidence? Explore Rear.
Note
These projects are early-stage and under active development. APIs, extension points, and persisted state formats may change between releases.
We welcome researchers and builders working on agent harnesses, evaluations, coding-agent infrastructure, and continuous learning systems. Join the GEAR Discord to follow development and share what you are building.