Darwin-Agent/HarnessX

What it solves

Most agent frameworks focus on swapping the underlying model, but changing an agent's behavior (e.g., switching from a coding agent to a research agent) usually requires rewriting the agent's logic. HarnessX decouples the model from the "harness"—the pipeline of tools, memory, and processors that define how an agent behaves—allowing users to change behaviors without rewriting the agent.

How it works

It uses a "harness foundry" approach where agents are created by pairing a ModelConfig (routing and provider settings) with a HarnessConfig (the behavior pipeline). The behavior pipeline is composed of reusable processors across nine dimensions (such as context, control, evaluation, and memory) that can be combined using a pipe operator (|).

To improve performance, HarnessX employs two evolution loops:

1. Harness Evolution: A meta-harness analyzes agent trajectories to automatically search for optimal processor combinations and prompt strategies.
2. Model Evolution: Reward-annotated trajectories are fed into RL fine-tuning (via VERL) to improve the base model.

Who it’s for

• AI developers who want to rapidly prototype and iterate on agent behaviors without rewriting code.
• Researchers focusing on agent optimization and self-evolving systems.
• Teams needing to deploy agents across multiple platforms (via the included IM Gateway for Slack, Discord, etc.).

Highlights

• Composable Pipeline: A 9-dimension behavior pipeline using a modular processor system.
• Self-Evolution: Ability to automatically optimize harness configurations and co-evolve the model via RL.
• IM Gateway: Built-in support for connecting agents to Feishu, Telegram, Slack, Discord, and DingTalk.
• Harness Lab: A dedicated UI for managing and testing agents.
• Light-Memory: A custom file-based memory system with time-decay and git versioning.