Harness adapters: Safe RL for efficient agentic development

Harness adapters are small LoRA modules trained with reinforcement learning separately for agent harnesses such as pi, opencode, and hermes, allowing a shared base model to be specialized for each harness’s tools, prompts, and control flow without modifying the harness itself. The project reports that early held-out evaluations showed fewer tool calls while maintaining or improving accuracy, including for TwIL-8B, a reasoning model not previously trained for agent use. Training combines data-analysis tasks with Repo2RLEnvs, verifiable coding environments reconstructed from shared agent trajectories and categorized mainly as build tasks, though the authors note the current dataset is heavily skewed toward easy build scenarios. To reduce reward hacking, every tool call is routed through safeagents, a rule-based proxy that blocks actions such as editing tests, accessing answer sources, bypassing proxies, or seeking credentials, assigning blocked rollouts zero reward. The system uses asynchronous GRPO training, shared vLLM serving, and adapter-based checkpoint evaluation to reduce compute costs, while publishing environments, rules, training configurations, evaluations, and blocked calls through a live dashboard; the authors encourage developers to share sanitized trajectories and train adapters for additional agent systems.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.