AI Harness

The key point is not to simply give an agent powerful tools, but to provide tools that are carefully designed to match the model’s inherent capabilities. As concrete evidence, this presents the three-stage evolution of the AskUserQuestion tool. Early attempts tried adding parameters to ExitPlanTool or parsing specific Markdown formats, but ultimately converged on a dedicated tool-calling interface that the model can understand most easily and respond to most clearly. It also emphasizes that tools should evolve as model performance improves, illustrated by the transition from a simple to-do list (TodoWrite) to a Task Tool that supports inter-agent collaboration and dependency management. It points out the limitations of RAG approaches that inject all context up front, and explains the importance of Progressive Disclosure, where the agent uses tools such as Grep to discover and build its own context—maximizing performance by filtering out irrelevant information and focusing on what matters.