Chat is a conversation. Agents are a control loop.

Part 3 added tools. This part is why that is not “chat with plugins.” The shape of the computation changes. Errors stop being local to one paragraph. They travel.

Traditional generative AI

One shot. You speak. It answers. Done.

User → Prompt → Model → Answer

If the answer is wrong, you prompt again. Nothing else moved unless you copied it somewhere.

An AI agent

A goal, then a repeating inner loop: reason, plan, use a tool, observe what happened, and go again until a result (or a stop).

User
  ↓
Goal
  ↓
Agent
  ↓
Reason → Plan → Tool → Observe
           ↑           ↓
           └── Repeat ─┘
                 ↓
               Result

The difference is the observe → repeat cycle. The system is not finished when it produces words. It is finished when the world (a ticket, a file, a calendar, a test suite) looks closer to the goal — or when a human or a policy cuts the loop.

Researchers named a close cousin of this pattern ReAct: interleave reasoning traces with actions, then take in environment feedback. Google’s ReAct note is the short public explainer. You do not need the paper to use agents. You need the diagram. Thought without action is chat. Action without observation is a script that cannot notice it failed.

Why the loop changes the risk

Each tool call can fail: timeout, wrong endpoint, empty result, permission denied. The next plan is built on that observation. A wrong summary wastes five minutes. A wrong observation can send the next call to the wrong customer.

Permissions matter because the badge is reused every turn. Logging matters because “what did it say?” is the wrong audit question. The right one is “what did it do, in what order?”

You would not hand a new hire every password on day one. Keep that instinct. The loop is junior staff with a badge, not a smarter autocomplete.

Example

Traditional: “Summarize this PDF.” One answer. You file it.

Agent: “File a support summary from this PDF and the last three tickets.” Search, read, draft, attach, submit. If ticket #2 was the wrong account, the submit step inherits the mistake.

Same model family. Different machine.

Conclusion

Once you see the loop, the rest of the series is easier. Part 5 is what goes wrong inside it. Parts 7–9 are how enterprises specify, test, and ship it.

Takeaway: Traditional AI is prompt → answer. Agents are goal → reason → tool → observe → repeat.

Sources

Part 3: The Agent Era
Part 5: The Uncomfortable Reality