Guide · Handing whole tasks to AI

A chatbot gives you something to action. An agent takes the action.

For anyone whose feed keeps shouting about agents without ever saying what one is.

People search for agents expecting magic or menace. The working definition is calmer and more useful: you give an agent a goal, some background and the tools it needs, then you review its work. That is exactly what you would give a new staff member, which means you already know how to work with one: like a manager.

The idea

You give an agent the goal, not the steps. Ask for every PDF on a website and it will try one technique, fail, try another, and keep going until the job is done. You set the outcome and the boundaries; it works out the route.

Which makes the shift managerial rather than technical. You move from making the work to editing it, so the skills that matter are the ones a good manager already has: knowing what good looks like, spotting what is wrong, and asking for the change in a way that lands.

Start smaller than the hype suggests. Take one task you would otherwise hand to a person, work out what the best person for that job would need to know, and configure exactly that. One agent, one job.

A chatbot hands you a draft. An agent does the job. Give it the goal and the tools it is allowed to touch. It picks the route, uses what it needs, and comes back with work for you to review.

Stand in the loop early, not only at the end. One check at the end meets everything at once, after it has all compounded. Two check-ins along the way cost minutes and save the rebuild.

How it works

  1. IntelligenceThe underlying model: the brain. You rent this part, and you can swap it as models improve.
  2. ContextWhat it knows about you and your organisation: the onboarding. An agent without context is an intern on their first day, sent off with no induction.
  3. ToolsWhat it can actually touch: email, calendar, files, the web. Picture a body around the brain: the tools are the limbs, and they move only with the permissions you grant.
  4. SkillsIts position descriptions: how your organisation does each task, written down once so the agent does it your way every time.
  5. ScheduleWhen it runs: every morning, every Friday, when something arrives. The schedule is what turns a chat into a colleague.

The parts

The five-part anatomy of an agent. Step through the parts, then take the design prompt at the end.

  1. IntelligenceThe model underneath: the brain. It already went to every university and read the whole internet. That part is solved, and it is the part you rent.The intelligence is the least interesting part. Everyone has the same brains available.
  2. ContextEverything it knows about you: your organisation, your voice, your priorities. This is the onboarding, and it is the part most agents never get, which is why most agents disappoint.An unonboarded agent is an intern sent to work on day one with no induction.
  3. ToolsThe limbs: what it is allowed to touch. Email, files, calendar, the web. Grant tools the way you would grant a new colleague system access: deliberately, and least first.Without tools, it's just a chatbot.
  4. SkillsThe position descriptions. A skill states how your organisation does one task, so the agent does it your way, and so does the next agent, and so does the person who joins next month.
  5. ScheduleHow often it runs. Every afternoon, sweep the inbox. Every Friday, draft the report. The schedule is the difference between a tool you remember to use and a colleague who shows up.
    6. The prompt for schedule
    I want to offload one task to an agent. Interview me, one question at a time, about how I do this task today, what good looks like, and what it would need to know. Then propose the agent: its context, its tools, its skill written as a position description, and its schedule. Flag anything in this task that should stay with a human.