HarnessX

HarnessX is a Python framework for developers building AI agents whose behavior they can reuse and adapt. It separates model choice from tool use, memory and context handling, so changing an agent's role doesn't require rewriting it.

• Composable behavior: Combine reusable modules for tools, memory, context processing, tracing and sandbox execution into different agent configurations.
• Model routing: Assign models to different roles and configure fallback providers independently of the agent's behavior.
• Behavior optimization: Analyze agent runs and search for better prompts, module combinations and tool configurations while keeping the model fixed.
• Training integration: Use reward-annotated agent runs for supervised fine-tuning and reinforcement learning, with VERL integration for model training.
• Messaging connections: An optional gateway connects agents to Feishu, Telegram, Slack, Discord and DingTalk, with a browser console for managing channels and sessions.

The open-source project uses the MIT license. It provides a Python SDK and a locally installed Harness Lab frontend. Model configurations used in its benchmarks include Qwen 3.5 9B and GPT-5; the framework keeps provider selection separate from the behavior modules.