# Harness-RL

A reinforcement-learning framework that trains language-model agents inside a central-agent multi-agent harness without flattening execution into a single token sequence, routing gradients to action and argument tokens separately.

Harness-RL constructs black-box trajectories from interface calls and session prefix trees, then applies Conflict-Aware Policy Optimization (CAPO) to route gradients from action tokens and argument tokens to their corresponding parameter subspaces. It supports both central-only training and joint multi-agent training, and was accepted at PCC 2026.

StackPlanner provides the production agent harness used for rollouts, while a companion component, slime, provides distributed rollout, SGLang serving, Megatron training, checkpointing and weight synchronization. It targets Linux x86_64 with NVIDIA GPUs, uses the Search-R1 retriever through StackPlanner's web_search interface, and includes make targets for environment setup, data preparation and retrieval configuration.

Website: https://github.com/jiangxinke/Harness-RL