There is a strain of agent demo, beloved of conference keynotes, in which a single ambitious task is completed end-to-end by a single dazzling agent. The task is usually something like booking a holiday or building a small application. The demo always succeeds. It is also, almost without exception, not the way agents are actually deployed in working organisations.

What is happening in practice is considerably less photogenic. Companies are taking processes that used to be measured in human-weeks (compliance reviews, invoice reconciliation, technical documentation, customer-issue triage) and rebuilding them as small constellations of narrow agents, each with a single job, monitored by a workflow that knows what to do when one of them fails.

The numbers, where I have been able to see them, are remarkable in a quiet way. Cycle times for the most heavily-instrumented processes have fallen by factors of three to ten. Error rates are roughly comparable to the human baseline, slightly worse in some categories and slightly better in others. Costs are lower, though not dramatically so, because the human time saved has been partially redirected into the supervision of the agents.

This is not the future the demos promised. It is, in many ways, more interesting than the future the demos promised. The first useful agents are not turning out to be brilliant generalists. They are turning out to be reliable specialists, deployed in systems that take their reliability seriously enough to monitor, retrain, and replace them. Which is, when one thinks about it, how most useful technologies have always worked.

The Boring, Important Work of Putting Agents to Work