Summary

HarnessX provides a foundry for creating, adapting, and evolving AI agent harnesses, decoupling agent behavior from the model. It enables behavior swapping (e.g., coding to research) without agent rewrites, targeting developers and researchers seeking flexible, high-performance AI agents via adaptable and evolutionary capabilities.

How It Works

The core innovation separates agent behavior configuration (the harness) from the base model. HarnessX composes agent behaviors via a 9-dimension pipeline of reusable "processors" combined with |. The system observes performance, automatically searches optimal harness configurations ("Adapt"), and uses reward-annotated trajectories for SFT/RL training ("Evolve"), reducing the cost of switching agent behaviors or adding features.

Quick Start & Requirements

Installation uses a one-click curl script (bash install.sh) or non-interactive --all, installing uv, Python 3.12, harnessx, and optionally Node.js for the Harness Lab frontend. Manual install uses uv. A CLI runs agents, loads configs, resumes sessions, and launches the Harness Lab UI (hx lab). An optional IM Gateway connects agents to platforms like Slack/Discord.

• Primary install: curl -sSf https://raw.githubusercontent.com/Darwin-Agent/HarnessX/main/scripts/install.sh | bash
• Prerequisites: Python 3.12, uv (installed by script), Node.js (optional).
• Links: Architecture Docs, Gateway Docs, Benchmarks, Roadmap.

Highlighted Details

• Composable Behavior Pipeline: 9-dimension pipeline allows flexible composition of agent behaviors using reusable processors.
• Harness Evolution: Auto-searches optimal harness configurations, showing significant performance gains (e.g., +14pp on GAIA benchmark with Qwen 3.5 9B, 62% to 84% with GPT-5).
• Model-Harness Co-Evolution: Compounding improvements via evolving both harness and model through RL fine-tuning.
• Harness Lab UI: Integrated React-based UI for managing agents and configurations.
• IM Gateway: Facilitates agent connectivity to messaging platforms.

Maintenance & Community

Developed by the "Darwin Agent Team," HarnessX is open-source. Contributions are welcomed for processors, memory backends, benchmarks, RL recipes, and Lab UI improvements, guided by CONTRIBUTING.md. Community channels like Discord/Slack are not explicitly detailed.

Licensing & Compatibility

Released under the permissive MIT License, allowing broad compatibility for commercial use and integration within closed-source projects.

Limitations & Caveats

Advanced features like Bayesian Optimization for harness search, a community marketplace (HarnessHUB), multimodal memory, and closed-loop Harness Memory Evolution are under active development (roadmap items marked [ ]). Some sophisticated self-evolution and advanced memory capabilities are planned for future releases.