single mention · post-training
Unified white-box RL environment
A unified white-box RL environment represents an agent harness as a collection of configurable, composable modules such as tool interfaces, system prompts, context management strategies, skills, memories, and subagents.1 By composing these modules through configuration, it can instantiate mainstream harnesses and generate diverse harness configurations during RL training, preventing the model from overfitting to a single harness's conventions.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
represents an agent harness as a collection of configurable, composable modules, including tool interfaces, system prompts, context management strategies, skills, memories, subagents, and other components
— k3 tech reportComposing these modules through configuration, the environment can instantiate mainstream harnesses
— k3 tech report
Filed under post-training › agentic post-training