HarnessX

HarnessX is a self-hosted Python framework for developers and researchers building AI agents for coding, research, or assistant tasks. It separates model selection from agent behavior, so you can change memory, tools, or safety checks without rewriting the agent. It's open source under MIT.

Reusable processors and prebuilt bundles define how an agent handles context, retrieves memories, evaluates results, and recovers from errors. Model routing supports fallback providers and different models for different roles. The framework includes tracing, checkpoints, and metrics for inspecting runs.

HarnessX can analyze past runs and search for better combinations of prompts, tools, and processors while keeping the model unchanged. It also records runs with reward annotations for supervised fine-tuning and reinforcement learning, with a VERL integration for model training. Its benchmark examples use Qwen 3.5 9B and GPT-5.

You can work through a command-line interface, Python SDK, or the browser-based Harness Lab running on localhost. The framework supports local and Docker sandboxes as well as E2B. Model calls depend on the provider you choose: the Anthropic integration sends requests to its cloud API and requires an API key, so running the framework locally doesn't make those calls offline.

An optional messaging gateway connects agents to Feishu, Telegram, Slack, Discord, and DingTalk. Its web console manages channels, sessions, and workspaces.