ChatGPT Work, Claude Code: what counts as an AI agent

The explainer's framing matters because much of the confusion around AI agents comes from lumping ordinary chatbots together with tools that actually act. By requiring that an agent design its own steps, operate tools like browsers and files, and verify and retry its own output, the piece draws a line that maps onto how vendors now describe their products — Anthropic's documentation for Claude Code, for instance, describes reading a codebase, editing files, running commands and checking results, with permission requested before changes by default.

The six business examples follow a consistent pattern: the agent gathers and organizes, and a person keeps the judgment. In sales, the agent researches a prospect and drafts an email, but a person checks the recipient and content before sending. In accounting, it reads invoices and flags mismatches against purchase orders, but payment approval stays with a person. The same split appears in support, HR and code review, which is why the piece treats review design, not model capability, as the deciding factor.

The caveats section extends that logic to risk. Claude Cowork is described as reading and writing only within folders the user connects and asking permission before deleting files, an example of limiting scope by design. The advice to start with research and document creation — marketing's competitor research is the named example — reflects the same caution, since the output can be checked by eye and errors carry little consequence.