跳到正文
DEV Community· SmashingApps.com·· 16 小時前AI 評分55

別把所有聊天機械人都稱為智能體:真正智能體的目標、規劃、行動與迭代循環

Stop Calling Every Chatbot an Agent: The Goal Plan Act Loop That Actually Earns the Name

AI 導讀

文章指出,聊天機械人只按提示詞回覆,智能體則接受目標、拆解步驟、呼叫工具、評估結果並反覆迭代,直至完成目標或停止。文中以是否具備這個循環區分工作流與智能體,並建議固定步驟或單次 LLM 呼叫已足夠時,不必加入智能體;若授予工具權限,應設最小權限、人工審核及操作記錄。

正文

Stop Calling Every Chatbot an Agent: The Goal → Plan → Act Loop That Actually Earns the Name

Short answer: a chatbot takes a prompt and returns an answer. An agent takes a goal, breaks it into steps, calls tools, checks its own results, and keeps going until the goal is done — or stops. That loop is Goal → Plan → Act → Evaluate → Iterate, and if your system does not have it, it is not an agent, no matter what the marketing page says.

According to SmashingApps.com, generative AI creates content — text, images, audio — from prompts. Agentic AI goes further: it sets goals, plans multi-step actions, uses tools, and takes semi-autonomous actions to achieve outcomes. The difference is not the model. It is the loop around the model.

The two interaction models, side by side

Generative AI runs on Request → Response. You ask, it answers, the conversation ends. That is enough for drafting, summarising, and brainstorming, and it is where most "AI features" still live.

Agentic AI runs on Goal → Plan → Act → Evaluate → Iterate. You set a goal. A planner breaks it into tasks. An executor calls tools — search, APIs, a database. An evaluator checks the results and decides what to do next. The loop repeats until the goal is reached or the system stops. That is the whole definition, and it is testable: if you remove the loop, does the system still do the job? If yes, it was never an agent.

What an agent is actually made of

An agent is built by combining generative capabilities with orchestration logic, memory, and tool access. In practice that means three parts: an LLM for planning and reasoning, a tool registry it can call, and an execution controller that runs the loop and keeps a log.

Anthropic draws an important architectural distinction in its guide to Building effective AI agents: workflows are systems where LLMs and tools are orchestrated through predefined code paths, while agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. Both are agentic systems. Only the second one is an agent in the strict sense.

And here is the part most teams skip: Anthropic found, working with dozens of teams, that the most successful implementations were not using complex frameworks or specialised libraries. They were building with simple, composable patterns. Start with the simplest solution possible, and only increase complexity when needed. For many applications, a single LLM call with retrieval and in-context examples is enough — no agent required.

Two examples that pass the test

First, a research agent. Give it a goal: find three recent blog posts on remote-work productivity with their key findings. It searches the web for competitor pricing and posts, compiles a spreadsheet, emails the results to you, and schedules a follow-up meeting. No human touched the middle steps. The loop did the work.

Second, an operations automation. It monitors servers, triages alerts, attempts a safe restart, and notifies the on-call engineer if the problem persists. Again, the value is not the model writing a status message. It is the system acting, checking whether the restart worked, and escalating when it did not.

If your "agent" cannot show you that loop — goal, plan, tool calls, evaluation, iteration — it is a chatbot with a workflow diagram.

When NOT to build an agent

Agentic systems trade latency and cost for better task performance, so the tradeoff has to make sense. Ask three questions before you build one.

First, is the task actually multi-step and unpredictable? If the steps are fixed and known in advance, a workflow with predefined code paths is cheaper, faster, and easier to debug. Use a workflow.

Second, can a single LLM call with retrieval do it? If yes, do that. Adding an agent loop to a one-shot task adds failure modes for no benefit.

Third, what happens when it goes wrong? An agent that can send emails, run code, or touch production data needs guardrails before it needs features. If you cannot answer that question, you are not ready to give it tools.

The guardrails that matter

If you do give a system agency — let it act — these are the non-negotiables, in order.

Give it the least privilege possible: only the tools and permissions it absolutely needs. Nothing else.

Keep a human in the loop: require explicit review for high-impact actions — payments, deletions, legal messages, anything you cannot undo.

Log everything: plans, tool calls, and decisions, with an explainable transcript you can audit later.

Add rate limiting and kill switches, so a runaway loop stops instead of burning your budget or spamming your users.

Validate and sandbox: test agents in isolated environments before production, validate and sanitise any data they use when calling external systems, and start with manual review for the first 100 runs before switching to more autonomy.

The multi-agent footnote

There is one more pattern worth knowing, because it is where the token bill explodes. In a multi-agent system, a lead agent plans the work and subagents explore different parts of it in parallel, each with its own context window, before condensing the findings back. Anthropic's multi-agent research system works this way, and it outperformed a single-agent setup by a wide margin on its internal research eval — but multi-agent systems used about 15 times more tokens than a standard chat interaction. That is a fair trade for a hard research problem. It is a terrible trade for a task a single call could have done.

The bottom line

Generative AI answers. Agentic AI acts, checks, and iterates. The test is simple: show me the loop. Goal, plan, tools, evaluation, iteration — with logging and a human gate on anything irreversible. Build the simplest thing that passes that test, and stop calling the rest agents.

For the full comparison, including the design patterns, the safety checklist, and how to choose between the two, see the complete guide to agentic AI vs generative AI on SmashingApps.com.

來源:DEV Community · dev.to