By function: the recurring shapes
Across industries the same shapes recur. Software engineering: agents that reproduce a bug, write the fix, run the tests, and open the pull request. Support operations: ticket triage, account lookups, drafting and — with approval — sending resolutions. Research and analysis: multi-source gathering, reconciliation, and a structured brief at the end. Back-office operations: reconciling records across systems that almost agree, chasing exceptions that rule-based automation kicks out. IT and security operations: enriching alerts, correlating across tools, proposing containment steps for human sign-off.
What separates these from a chatbot with extras
Run any candidate through three questions. Does it carry a goal across multiple steps without re-prompting? Does it act through tools — writing to systems, not just producing text? Does the outcome depend on decisions the model made along the way? A drafting assistant fails the second test; a scripted workflow with one model call fails the third. The label on the product does not settle it — plenty of things marketed as agents are single-pass generative tools, and the distinction is worth keeping because the oversight each needs is different.
The readiness lens on every example
Each example above is also a list of permissions. The coding agent holds repository write access; the support agent touches customer records and outbound mail; the operations agent reaches systems of record. Reading examples this way — what does it write to, how far does its loop run unattended — is more useful than collecting use cases, because it converts inspiration into scoping: the same questions that pick your first agent also size the controls it ships with. Start narrow, reversible, and verifiable; the build guide turns that into a sequence.