The orchestration layer around a language model that equips it with tools, memory, and control flow — turning a text predictor into an agent that can plan, act, and loop.
An agentic harness is the infrastructure wrapped around a language model that transforms it from a single-turn text predictor into an autonomous agent. The model itself remains unchanged; the harness provides the tools it can call (web search, code execution, database queries, API calls), the memory systems that persist state across turns, the routing logic that decides which tool to invoke, and the loop that runs the model repeatedly until a task is complete.
Without a harness, a model answers one prompt and stops. With a harness, it can plan a multi-step research task, execute each step, observe the results, and continue until it has a complete answer — or until it decides to ask the user for clarification.
Tool definitions — structured descriptions of each available capability (search, browse a URL, run Python, call an API). The model reads these definitions and decides when and how to invoke them.
Tool executor — the code that actually runs a tool when the model requests it and returns the result back into the context window.
Memory — short-term (what has happened in this conversation), working (scratchpad notes the model writes to itself), and long-term (a vector store of facts persisted across sessions).
Loop controller — the orchestration logic that feeds tool results back to the model, checks whether the task is done, and either completes or iterates.
Google's AI Mode is a harness: Gemini sits at the centre, with search, Python, time, location, and URL browsing tools wired around it. When a user asks a research question, the harness routes sub-queries through the search tool, retrieves grounding snippets, feeds them back to Gemini, and synthesises a final answer. DEJAN's query fan-out research documents exactly this pattern. The multi-step research agent pattern — where one agent spawns sub-agents for parallel research — is an extension of the same harness idea.
Understanding the harness explains behaviours that seem mysterious if you think of the model alone. Why does the model sometimes search and sometimes not? The harness's routing logic decides. Why do AI answers cite certain pages and not others? The harness's retrieval step selected them. Optimising for AI search means optimising for the whole harness — being retrievable by the search tool, survining the snippet extraction, and surviving the model's synthesis step.
