From Chatbots to AI Agents

An answer becomes a working loop

The word “agent” has no single definition used everywhere. Anthropic distinguishes fixed workflows from systems in which the model directs its own process and tool use. A typical agent can act, inspect what happened and decide what to do next. That is more than producing one response. ^[1]

In our hypothetical trip task, the system first establishes dates and constraints, checks available information, compares options, then revises its plan if a suitable connection cannot be found. This is an illustrative workflow, not a promise that every agent can reliably arrange travel.

Tools turn language into action

A model alone cannot update your calendar or submit a form in another service. It needs access to tools and an environment where those tools work. Anthropic describes the model, its operating instructions, tools and environment as distinct parts of an agent system. Changing the permissions changes what the system can do. ^[2]

That is why comparing agents only by the fluency of their answers misses something important. A beautifully written itinerary could contain a nonexistent connection. A useful system needs to check the underlying results, not just explain its plan confidently.

Autonomy needs boundaries

Errors can accumulate across a sequence of actions. An agent may misunderstand the goal or encounter malicious instructions hidden in material it reads—a problem called prompt injection. Tool access, permission limits and checks all contribute to managing these risks. ^[1][2]

For our trip example, researching options and preparing a comparison are different from paying for tickets. You can define a boundary: assemble the options, then stop before a purchase. That makes the task useful while keeping an important decision explicit.

The interesting question is therefore changing from “Can AI answer this?” to “Which parts of this job can it complete, check and hand back reliably?”