Training the Whole Agent, Not Just the Model

One-line takeaway

Tencent's HarnessAdapt-RL adapts the coding agent's harness during RL — one teacher writes task-specific repair skills, another evolves tools and hooks kept only if pass rates improve — so the whole agent is trained, not just the model.

腾讯的 HarnessAdapt-RL 训练编程 agent 时同时改进框架：冻结教师为错误修复策略生成任务级技能，另一教师演化工具和钩子，只保留能提升通过率的改动。