自主智能体
定义
自主智能体在有限的人类输入下在较长时间范围内追求目标。 They plan, use tools, and adapt when the environment or task changes (例如 coding agents, research assistants).
They sit at the “high autonomy” end of the agents spectrum: 不是一个用户轮次一个响应，而是运行长循环 (plan → act → observe → replan) until the goal is met or a limit is hit. Subagents and 推理 patterns (例如 ReAct, ToT) are often used inside autonomous agents to structure planning and action.
工作原理
The agent starts from a goal (例如 “implement feature X”). It plans (possibly breaking into steps or sub-tasks), then acts (工具调用, code edits, search). The observe step captures results (tool outputs, errors, state) and feeds back into plan for the next iteration. The loop combines planning, memory (what was tried, what worked), tool use, and often reflection (例如 self-critique). It runs until a stopping condition: task done, step/budget limit, or human-in-the-loop check. Safety and oversight (例如 approval gates, rollback) are important when autonomy is high.
应用场景
自主智能体适用于长期、多步骤的工作，系统必须在没有逐步人工输入的情况下进行规划、行动和适应。
- Long-horizon coding agents that plan, edit, and test
- Research assistants that gather sources, summarize, and iterate
- Data pipelines that adapt when inputs or schemas change