Agent demos are impressive and useless. What works in practice is small: one job, clear inputs, a limited set of actions, and a human who sees the output before anything irreversible happens.
1. The inbox triage agent
Reads incoming mail on a schedule, classifies it against your own categories, drafts replies for routine items, and flags anything that needs judgment. It never sends without you. Saves twenty minutes a day and, more importantly, removes the decision fatigue of a full inbox.
2. The meeting follow-up agent
Takes the transcript after each call, extracts commitments with owners and dates, and drops them into your task list as drafts. Guardrail: everything ambiguous goes in a "confirm" list rather than being guessed.
3. The research brief agent
Given a company or topic, it gathers public information, then writes a one-page brief in your fixed format with a source next to each claim. Guardrail: anything it cannot source gets labelled as unverified.
4. The document check agent
Watches a folder, and for each new document produces a short summary plus a list of what is missing or inconsistent against a checklist you wrote. Useful for proposals, contracts and reports.
5. The weekly digest agent
Collects the handful of sources you actually trust, summarises what changed, and answers one question: does anything here change what I should do next week. Guardrail: a hard limit on length, or you have invented another thing to read.
What they all share
One job. A narrow set of actions. A written definition of done. A log you can inspect. And a human approval step before anything leaves the building or costs money.
Where people go wrong
Giving one agent five jobs, letting it act without review, and having no record of what it did. Every agent failure I have seen traces back to one of those three.
Try this today
Name the one recurring task you would hand a careful assistant, then write the three rules that assistant must never break. That is your first agent specification.