Details
- Hugging Face and Adithya S. K. released a practical, open guide to training language-model agents across multiple agent harnesses.
- The guide addresses a major reproducibility issue: the same model and weights scored 62% in one harness but only 33% in another.
- Its approach uses a proxy between the harness and model rather than modifying the harness itself.
- The proxy supports four API formats used by coding agents: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Gemini.
- It records the token IDs and log probabilities sampled by vLLM, allowing those interactions to feed reinforcement-learning training.
- In reported experiments, Liquid AI’s LFM2.5-2.6B improved from 42% to 54% after training across four harnesses.
- A shorter-agent-trajectory reward reduced tool calls by 31%, while training only in OpenCode improved that harness from 34% to 58%.
- The project releases the capture proxy, trainer, tasks, supervised fine-tuning data, training code, and seven trained models openly.
Impact
The release targets a growing problem in agent development: performance depends heavily on the execution harness, not just model weights. By allowing one training workflow to operate across Claude Code, Codex, OpenCode, and other compatible interfaces without rewriting each harness, the approach could reduce engineering costs and make open-model agent training more reproducible. Its open components also give smaller labs a practical alternative to proprietary, platform-specific training pipelines.
