1分钟看图找到关键要点👇

大模型不是马。是大脑，而且是一颗刚刚觉醒的大脑。说起来，

一、Agent 不止是 model+loop

早期我理解的Agent 工程。可以被写成一段伪代码：

whilenot done:
reason
act
observe

使用者输入一句话，模型思考一下决定调用工具。工具返回结果，模型再思考，再决定接下来。

这也是 ReAct 的基本形态。Thought、Action、Observation 交替出现。模型在生成内容的过程中调用工具，再把工具结果放回接下来推理。

早期Agent概念刚出。我做了一个 demo，这个循环完全够用了。命令行里输出 token，中间插入 tool call。再把 observation 塞回 prompt，最终模型给出结论。

但最近开始深入去做Agent产品。过程中疯狂调研codex、lobehub、goose、opencode、PI、Flue等优秀Agent产品，发现远远不止这段伪代码所表达的。

真正的产品还会遇到一组运行时问题：

• 使用者刷新页面之后刚才的工具审批还在吗？
• 工具执行到一半，后端进程重启了下一次从哪里恢复？
• 子 Agent 在后台跑，主线程要显示什么？
• 生成了一个 artifact。内容本身不塞进上下文，那它的引用、状态、归属在哪里？
• 使用者点了 stop。哪些东西应该取消，哪些东西应该保留？

ReAct 不回答这些问题。

ReAct 解释的是模型怎么思考和行动。Agent 产品还需要一层工程程序：把模型做过的事，落成可恢复、可控制、可展示、可验证的软件事实。

下面我把这层程序叫做 Agent Harness。

二、ReAct 的解释边界

ReAct 的最小单元是：

Thought -> Action -> Observation

This abstract describes how models behave: think first,act next。observe results.

run.started
message.created
assistant.text.delta
tool.call.created
tool.approval_required
tool.call.running
tool.call.completed
artifact.created
run.finished

三、“事实从哪里来”探讨与设计原则

• When approval arrives late,a simple “yes/no” string is all you get;.
"
• Artifact references become mere strings mentioned in prompts;
"
• Stop button remains just a front‑end state toggle;. "
" "

";

Key Insight If you treat Harness purely as “model + loop”,you miss what truly matters: where facts originate during runtime.

Runtime responsibilities

html tool.approval_required // emitted immediately upon needing approval artifact.created // written when file/document materializes control.resume/run.stop // invoked via front‑end commands routed through runtime

These events anchor state transitions within running workflow.

四、“Loop → 协议”转化路线图

**

---

五、“State Schema”——定义事实边界

When building any Agent – wher work‑oriented or domain‑specific – start with a clear state schema. It determines which facts persist across page reloads / process restarts versus transient view data.

sql SELECT * FROM agents WHERE pendingApproval IS NOT NULL;

Questions you must answer upfront:

sql -- Which facts must survive restart?-- Which facts need cross‑endpoint consistency?-- Which facts belong solely to view logic?-- Which facts represent live interactions?-- Which facts derive cleanly from messages?-- Which facts become top‑level states?

Example: Approval is both a UI button and a recoverable pause point.

ini agent.state.pendingApproval = { id。runId,turnId,toolCallId,toolName,arguments,policy }

When users click “Approve,” send control:

css agent.control.resume({ approvalId,decision : "allow" });按理说，

Runtime resumes from checkpoint & writes updated state;frontend consumes only this authoritative state.

Artifact Representation

Artifacts rarely fit naturally inside message bodies: y could live anywhere – filesystem。database or object store – yet ir metadata must be stable.

bash artifactRef = { id,type,title,status,ownerRunId,createdByToolCallId,contentRef };

With this reference everything else follows: right side panel ➜ history ➜ recovery flows.

Sub-Agent Tracking

Sub‑agents shouldn’t be embedded solely in prose.

lua subagent = { id,parentRunId,parentTurnId,title,status,startedAt,completedAt,resultRef。error }

Frontends derive badges/progress indicators purely from this slice.

---

六、“UI只能消费事实不能补写事实”

The frontend handles layout/interaction/streaming presentation & view state only. Runtime owns fact emission.

Rule: Runtime writes fact ➜ UI consumes fact ➜ View renders control elements.

---

七、“程序要记得发生过什么”

Harness formalizes critical run-time facts: • Which tool call was initiated?老实说，• Where does it pause for approval?• Which run produced artifacts?• Which stop command targets which run?

Ensuring se truths persist guarantees: – Page refresh retains exact pending approvals. – Process restart resumes correctly. – User stop halts execution without lingering background calls. – Artifacts stay locatable & auditable. – Sub-agent completions surface correctly tied back to parent turns.

All such invariants hinge on having well‑defined fact ownership baked into runtime storage rar than piecemeal front-end heuristics.

---

八、“”

An Agent product’s true challenge isn’t simply crafting an LLM + tools loop—it’s delivering deterministic software behavior that survives real-world operations:

• Refresh without losing progress • Restart restores exact position • Explicit pauses/resumes • Traceable outcomes

That’s precisely what Agent Harness delivers— turning fleeting model actions into durable software realities. 一句话

---