Agents keep the language model, then add hands.

Chat replies. Reasoning plans more carefully. An agent tries to finish a job: read a database, call an API, browse the web, run code, click through an application.

That is why demos look magical. Research a vendor list. Fill a spreadsheet. File a support summary. Propose a pull request. The model is no longer only narrating. It is operating.

If you want the control narrative — what it may touch, what you must approve — read AI Right Now Part 6. This part is the mechanism.

Tool calling: the model chooses, the runtime executes

The technical beat is tool use (also called function calling). You register tools with names, descriptions, and JSON-shaped arguments — search_web, query_orders, run_sql, send_email. The model does not magically grow USB ports. It emits a structured request. Your application (or an agent runtime) runs the function, then feeds the result back as the next observation.

OpenAI’s tools guide is the plain version of that contract: the model selects a tool; your code performs the action.

So an agent can:

  • Read databases — if you expose a query tool with a connection you control
  • Call APIs — HTTP to tickets, calendars, CRM, payments
  • Browse the web — search or fetch pages you allow
  • Execute code — a sandbox, not your laptop’s entire disk unless you are reckless
  • Interact with applications — APIs first; later, computer-use that clicks real UIs

The unit of work stops being “a reply.” It becomes “progress toward a goal.”

Why this is a different product than chat

Chat failure mode: a wrong paragraph. Agent failure mode: a wrong side effect. Timeouts, wrong endpoints, stale docs, the wrong recipient on the invite — each tool call is a door.

That is not an argument against agents. It is why the next two parts exist: the loop is different, and mistakes compound.

Example

You say: “Pull last week’s failed payments, group by reason, and open a ticket for anything over a thousand dollars.”

An agent might query the payments table, summarize, then call the ticketing API. At the good end, you review three draft tickets. At the bad end, it pages the on-call for a known test card, or writes a ticket with a customer’s full PAN because the query tool was too wide.

The request was one sentence. The blast radius was a database plus a pager.

Conclusion

The agent era is tools plus a loop. We will draw that loop next. Keep Part 6 of AI Right Now nearby whenever a demo asks for credentials.

Takeaway: Agents are language models with tools — the model chooses an action; your runtime actually does it.

Sources

Part 2: The Reasoning Era
Part 4: Why Agents Are Different