Hugging Face releases open tools for training agents across harnesses
Same weights perform differently depending on the harness used
TL;DR:
- Model scores vary sharply across different agent harnesses
- New proxy grabs tokens and logprobs from major APIs without code changes
- LFM2.5-2.6B rose from 42% to 54% success after multi-harness training
- Full stack including seven models released under open license
Headline
Hugging Face put out an open method for running reinforcement learning across several agent harnesses at once.
Summary
The team noticed that identical weights can post very different scores depending on which harness they run in. That means the harness counts as part of the environment. Their proxy sits between the model and the usual APIs from OpenAI, Anthropic, and Gemini, pulling token IDs and logprobs without touching Claude, Codex, or OpenCode.
Liquid AI’s LFM2.5-2.6B climbed from 42% to 54% when trained on four harnesses and cut tool calls by 31%. The proxy, trainer, tasks, data, code, and seven trained models are all public.
Analysis
Training now has to consider the model and harness together because tool schemas, orchestration, and execution feedback all change results. Models trained on several harnesses handled new environments better than ones tuned on a single harness. OpenCode-only training lifted that harness from 34% to 58%, yet the multi-harness version improved everywhere. Imitation learning from Qwen3.8-27B rollouts underperformed letting the target model explore and get real task feedback.
Releasing the capture layer and training code gives open-model teams a direct way to train agents that either specialize or hold up across harnesses. That raises the bar for closed providers and makes cross-harness testing a standard check.
Impact Assessment
High technical impact for open AI research.