什么是 AI Agent Harness？

An AI agent harness is the software environment that surrounds an LLM to give it the tools and context needed to complete multi-step tasks. It turns your basic LLM calls to an operational system executing real world task. The model reasons through a prompt and decides the actions. The harness connects the agent it to the tools, systems, memory and execution environments needed to carry out those actions. In a AI harness Agents typically run in a loop: an LLM decides what to do, a tool executes, a model evaluates the results and then continues in that loop until the task is complete. Model(Reason)+ Harness(Action)= Agent (Output) Or Model (Brain) Agent (as the Body) and Harness as the hands/tools the body uses to finish a given task. You get the picture. The reason - act - observe loop This is at the core of many AI agents. Understanding this loop is critical to understanding how a harness works. Reason- The model reads everything in its context, including the task, relevant memory and previous results, then decides what action to take next. Act - The harness carries out that action by running a tool, executing code in a sandbox, calling an API or writing to storage. Observe- The harness captures the result and feeds it back to the model as new context. Repeat - The model uses that result to decide what to do next. The loop continues until the task is complete. The ReAct: Synergizing Reasoning and Acting in Language Models- is published in this paper and is an excellent read. Ok enough of "simile", let's jump to a concrete example Let's say we have a file in our local system "sales and vendor commission.txt" and we want to perform complex and kind of weird calculation on this. This is your final prompt - "I have a text file called 'sales and commission.txt' in the current directory - if it exists, extract the column with header 'Sales Person Name' , calculate it's word count. Also, if the word count is > 20 calculate the vendor commission which is 15% of column header Sales or if the word count is dict[str, Any]: try: tree = ast.parse(expression, mode="eval") result = _safe_eval(tree.body) return {"result": result} except Exception as e: return {"error": f"Could not evaluate expression: {e}"} def tool_get_current_time(utc_offset_hours: float = 0) -> dict[str, Any]: now = datetime.now(timezone.utc) if utc_offset_hours: from datetime import timedelta now = now + timedelta(hours=utc_offset_hours) return {"iso_timestamp": now.isoformat(), "utc_offset_hours": utc_offset_hours} def tool_word_count(text: str) -> dict[str, Any]: words = text.split() return {"word_count": len(words), "character_count": len(text)} Enter fullscreen mode Exit fullscreen mode Step 3 - System Prompt self.system_prompt = system_prompt or ( "You are a helpful assistant with access to tools. " "Use tools whenever they would give a more accurate or reliable answer " "than reasoning alone (e.g. always use the calculator for arithmetic). " "Once you have everything you need, give a clear, direct final answer." ) Enter fullscreen mode Exit fullscreen mode Step 4 - This is the heart of the whole logic, Use the stop_reason to check if the Agent is done is the reason is tool_use then keep going in a loop if response.stop_reason != "tool_use": # Final answer reached. final_text = "".join( block.text for block in response.content if block.type == "text" ) return final_text # Otherwise, execute every requested tool call and collect results. tool_results = [] for block in response.content: if block.type != "tool_use": continue self._log(f"[tool call] {block.name}({json.dumps(block.input)})") result_json = self._execute_tool(block.name, block.input) self._log(f"[tool result] {result_json}") tool_results.append( { "type": "tool_result", "tool_use_id": block.id, "content": result_json, } ) Enter fullscreen mode Exit fullscreen mode Core loop: Send the conversation + tool schemas to Claude. If stop_reason == "tool_use", execute each requested tool locally and append the results as tool_result blocks. Loop back to step 1. Stop when Claude returns a final text answer (stop_reason == "end_turn") or max_iterations is hit (safety limit).