# Agent Harnesses and Agentic Search ## Harness The software around a model that turns it into an agent. The model produces text; the harness decides what to do with it. A harness typically handles: - **Tools:** letting the model run code, read files, call APIs, browse. - **Context management:** choosing what goes into the model's limited context window, summarising history, retrieving relevant files. - **The loop:** plan, act, observe the result, try again, stop when done. - **Guardrails:** permissions, sandboxing, cost limits, human approval steps. Coding tools such as Cursor, Claude Code and Cognition's Devin are harnesses wrapped around one or more models. Much of the product difference between them lives in the harness rather than the model, which is why a harness can swap in a cheaper open-weight base without users noticing. See [[GTM-Harness-State-of-the-Art]] and [[AI Agents Stack]]. ## Agentic search Search performed by an agent rather than a single query. The model issues a query, reads results, reformulates, follows links or code references, and repeats until it has what it needs. Used in deep-research products and in coding agents navigating large codebases. It is expensive in tokens but far better on hard questions than one-shot retrieval. ## Why they matter in the chessboard Harnesses and agentic search are a newer layer where engineers on both sides now build on each other's work, alongside [[API Distillation]], [[Open Weights]] and data labelling. They also generate large volumes of trajectories that feed back into training. Related: [[AI agents]], [[Autonomous Agents]], [[AI coding tool]], [[Embracing the Agentic Engineering Era]]