How an AI agent actually works, in five pictures
The version I wish someone had drawn for me before I started building these for health and wellness practices.
Most people picture a chatbot. You type, it types back, and nothing in the world changes. An agent is the other thing. It reads a situation, decides what to do about it, does it in real software, and then tells you what it did. The typing is the smallest part.
It runs in a loop
An agent is not one answer. It is a small loop that keeps going until the job is done or until it gets stuck. Read what came in. Decide what to do. Do it. Check whether it worked. Say what happened. Then round again with whatever is left.
The last two steps are where the value is. Anything can do something. Checking the result and admitting when it went wrong is what makes an agent safe to leave running.
What one is made of
Four parts, and only four. Instructions are what you want and how you want it done. Tools are the things it is allowed to touch: a calendar, a form, a message queue. Memory is what it knows about this person and this practice. Limits are the edges, the things it may never do without asking.
Where it fits in a practice
Not the care. Never the care. Agents belong on the repetitive edges of a week, the jobs that are the same every time and that nobody enjoys.
- Turning a new enquiry into a booked appointment
- Chasing the intake form somebody started and abandoned
- Reminding, rescheduling, and following up after a visit
- Writing the boring half of a record so a person can write the half that matters
Each of those is one narrow job with a clear finish line. That is the shape that works. An agent asked to run the front desk in general will run it badly.
The dial that decides everything
Every agent has one setting: how much it is allowed to decide on its own. Turn it up and it saves more time. Turn it up further and you start spending your evenings undoing what it decided. The two lines cross somewhere, and the whole job is finding where.
The agents that last do one narrow thing, write down what they did, and hand control back the moment they are unsure.
What stays human
Before I build anything, I sort every task in a practice into three piles.
If you want to start
Pick the single job that annoys you most this week. Make sure it is reversible and that you would notice if it went wrong. Give the agent that one job, make it write down everything it did, and read the log for two weeks. Then decide whether it gets a second job.
The agent that survives is boring. It does one thing, leaves a trail, and hands back control the moment it is unsure. Nobody writes a case study about that, which is exactly why it keeps working after you stop watching.