AI Agent · From One Reply to Controlled Action

01 · 会回答，等于会办事吗 01 · Is answering the same as doing?

让聊天模型说“我查过库存了”，和系统真的查库、读回结果、更新任务，有什么不同？ What separates a model saying “I checked inventory” from a system that actually queries, reads the result, and updates the task?

02 Agent 组件台：一次点亮一条真实消息流 Agent Workbench: Light One Real Message Flow at a Time

四块组合 four-part system

消息进度 Message progress

0/6

等待目标 waiting for goal

真实工具调用 Real tool call

read_inventory

任务状态 Task state

NEW

external state

03 Agent 不是一个气泡，而是 4 块系统 An Agent Is Four System Parts, Not One Bubble

职责分离 separation of duties

MODEL

模型选择下一步 Model chooses next step

根据目标与当前状态提出动作，但不直接冒充外部执行结果。 Proposes an action from goal and state, but does not impersonate an external result.

STATE

状态记录进度 State records progress

保存已知事实、当前步骤与待完成项，让任务不靠模型“似乎记得”。 Stores known facts, current step, and remaining work instead of relying on what the model seems to remember.

TOOL

工具执行动作 Tool executes action

按结构化请求读取信息或改变状态，必须受权限、参数与确认门控制。 Reads or changes state through structured calls governed by permission, parameter, and confirmation gates.

ENVIRONMENT

环境交回事实 Environment returns facts

库存、订单、文件等真实系统返回结果；它可能与模型预期不同。 The real inventory, order, or file system returns results that may differ from the model’s expectation.

04 最小机制：动作之后必须读回环境 Minimal Mechanism: Read the Environment After Acting

state_t+1 = update(result)

s_t+1 = update(s_t, result_tool)

只有工具真实结果进入 update，任务状态才从 NEW 变为 VERIFIED；模型说“完成”不触发状态更新。 Only a real tool result moves task state from NEW to VERIFIED. A model merely saying “done” does not update state.

模型提议动作 Model proposes action read_inventory(A-17) read_inventory(A-17)

工具访问环境 Tool reaches environment 返回 stock = 3 returns stock = 3

状态更新进度 State updates progress VERIFIED · 可发货 VERIFIED · can ship

05 边界实验：只留下会说话的模型 Boundary Test: Leave Only the Talking Model

主动制造失败 Create a failure

跳过工具与环境，直接写“已发货” Skip tool and environment, then write “shipped”

模型能生成听起来完成的句子，但没有真实调用、结果或状态更新。这是一次回答，不是完成任务的 Agent。 The model can generate a completion-sounding sentence, but no call, result, or state update exists. That is an answer, not an Agent completing work.

尚未执行。任务仍等待真实消息流。 Not run yet. The task still awaits the real message flow.

✗ 回复很长，所以它是 Agent。 A long reply makes it an Agent. → ✓ Agent 需要目标、状态、工具与环境反馈。 An Agent needs goal, state, tools, and environmental feedback.

✗ 模型说“调用成功”就算工具结果。 The model saying “call succeeded” counts as a tool result. → ✓ 真实返回必须由执行环境送回。 The execution environment must return the actual result.

✗ 有模型和工具就足够，不必记状态。 A model and tools are enough; state is optional. → ✓ 状态让进度、事实和未完成项可核验。 State makes progress, facts, and remaining work verifiable.

06 · 一句话带走 06 · One line to keep

Agent 是模型、状态、工具和环境围绕目标组成的行动系统；本页看清这四块，下一页再看它们怎样重复成循环并可靠停下。 An AI Agent combines model, state, tools, and environment around a goal. This lesson identifies the parts; the next shows how they repeat as a loop and stop reliably.