Everything so far has been how these systems are put together. Before we open up a real one, it is worth knowing where they are already running — because the answer is less futuristic and more ordi…
Everything so far has been how these systems are put together. Before we open up a real one, it is worth knowing where they are already running — because the answer is less futuristic and more ordinary than the coverage suggests.
Agents are quietly at work behind interfaces you have probably already used. Four areas where the value is clearest.
This is the largest category by volume. An agent reads a request, works out what it is, acts on it, and personalises the result.
That fourth point is the underrated one. An agent that hands over cleanly, with the context preserved, improves the human's job rather than competing with it. The version that answers badly four times and then dumps a confused customer on a colleague makes everything worse. Both are built from the same components; the difference is whether anybody designed the handover.
Agents help produce material quickly and consistently, which is why content teams adopted them early.
These rely heavily on stored instructions and memory to stay consistent across hundreds of outputs. Consistency at that volume is the part that is genuinely hard to do by hand, and it is where the time is saved.
In regulated or research-intensive fields, agents augment professionals rather than replacing them. The aim is speed with oversight.
Healthcare. Summarising patient charts and clinical notes, surfacing guideline-based options for a clinician to weigh, assisting with intake questions.
Finance. Monitoring transactions for anomalies, drafting summaries from market data, tracking regulatory change.
Research. Searching academic databases, extracting findings across thousands of documents, producing summaries with citations attached.
These lean heavily on retrieval and specialist data. They are also, not coincidentally, the tasks that pushed the major AI companies towards dividing work between several agents, because searching many sources at once splits up naturally. Chapter 11 shows what that looks like in a system you can use today.
Coding deserves separate treatment, because it is where these tools have moved fastest and where the working pattern is most distinctive.
An early coding assistant completed the line you were typing. A coding agent does something categorically different: it reads the codebase, forms a plan, edits several files, runs the tests, reads the failures and tries again. Claude Code and the agent modes in GitHub Copilot work this way.
What makes this work is a property software has and almost nothing else in this chapter does. Code can be checked automatically. It either compiles or it does not; the tests either pass or they do not. The agent can grade its own work without asking anybody.
Hold onto that, because it is going to matter a great deal in two chapters' time. A task where the system can tell whether it succeeded behaves completely differently from a task where only a person can judge.
Most people meet all of this through a single window: they type into ChatGPT or Claude and something happens. Sometimes an answer comes straight back. Sometimes the system goes away, looks things up, and returns with citations. Same box, very different machinery — and it never tells you which one you just got.