What AI agents are and how they differ from a chat with a model
The word "agent" gets stuck on everything, and the confusion gets in the way of deciding. The difference is one thing: a chat waits for you to open a window, an agent fires by itself. Below are the four parts it consists of, what genuinely gets automated, and what never gets given to one.
The word "agent" gets stuck on everything, and the confusion gets in the way of deciding. The difference is one thing: a chat waits for you to open a window, an agent fires by itself. Below are the four parts it consists of, what genuinely gets automated, and what never gets given to one.
---
"I already use AI, why would I want agents"
A fair question, and the confusion here is widespread. The word "agent" gets stuck on everything: on chatbots, on assistants, on ordinary automation flows.
The difference is simple and fundamental.
A chat answers a question. You came, asked, got an answer, left. The initiative is always yours. Nothing happens until you open the window.
An agent does work. It fires on an event — an email arrived, a document changed, a time came round — decides for itself what to do, and takes actions: writes, sends, fills in, notifies, calls other services.
The key difference is one word: initiative. A chat waits for you. An agent works when you are not there.
What an agent consists of
Four parts. Understanding that structure immediately eliminates half of all wrong expectations.
A trigger — what starts it. An incoming message, a new row in a spreadsheet, the time of day, a change of status, a client enquiry. With no trigger it is not an agent but the same chat.
Logic — how it decides. This is where a model works: parse the incoming item, understand what it is, choose a scenario. That part is what separates an agent from ordinary automation, where rigid rules decide everything.
Tools — what it can do. Send an email, write to a spreadsheet, create a document, call a service, notify a person. An agent with no tools only thinks.
Memory and context — what it knows. A knowledge base, history, your company's rules. An agent with no base answers in generalities, and that is the most frequent reason adoptions fail.
How this differs from ordinary automation
Automation existed before too: connections between services, "if — then" rules. It works reliably and costs little.
Automation cannot do one thing: handle the unstructured. It cannot read an email and work out whether it is a complaint or an order. It cannot decide what to do with an enquiry that fits no template.
An agent can. Hence the right division of labour: mechanical work to automation, work requiring understanding to an agent.
An error in either direction costs money. Putting a model where a simple rule would do is expensive and less reliable. Trying to parse living speech with rigid rules is pointless.
What genuinely gets automated
Not "everything", and it helps to understand the boundary.
Works well:
- Sorting incoming enquiries: what this is, who from, where to send it
- A first reply out of hours — with qualification rather than a brush-off
- Preparing drafts to a template from incoming data
- Collecting and consolidating information from several sources into a report
- Reminders about deadlines and statuses
- Tagging and categorising bulk material: reviews, enquiries, call recordings
Works with caveats: anything where the cost of an error is high. The agent prepares, a human checks and sends.
Never given to an agent: final decisions with legal or financial consequences, work with data that must not leave the building, and communication with a client on a large deal all the way to the end. The rule: a hot enquiry goes to a human.
The main mistake in adopting one
Automating chaos.
If a process is undescribed and everybody does it their own way, an agent will cement that state and make it harder to fix. The order is compulsory: describe first, automate second.
The second most frequent — an agent with no knowledge base. It answers in generalities, irritates people more than silence and buries the idea in the company for years.
The third — starting with the most important process. You start with something dull and your own: there a mistake costs nothing and experience appears.
Where people start in practice
Not with a big project. An order that works.
Step one — choose a dull task. One that repeats, irritates you and breaks nothing if it goes wrong. Sorting incoming items by subject, reminders, compiling the weekly report.
Step two — describe it in words. What happens, where it comes from, what has to come out, what the exceptions are. If you cannot describe it, it is too early to automate.
Step three — build the simplest version. One trigger, one action, no clever logic. Let it run for a week.
Step four — add understanding. Only where the rigid rules did not cope. Usually it turns out they coped in most cases.
Step five — put a human on the output. At first the agent prepares and you check. You can let it go unchecked once a month has passed with no error found.
That order takes weeks rather than months and, crucially, does not leave you with a system nobody trusts.
Where this gets built
There are several ready platforms and they differ in barrier to entry.
Visual flow builders — the lowest barrier, assembled with a mouse, covering most standard tasks.
Flexible automation platforms — more freedom, with the ability to write logic, suited to non-standard work.
Specialised agent environments — when you need complex scenarios with memory and many tools.
Starting with the first is sensible. Moving on when you hit a specific constraint rather than in advance.
How much this actually saves
Worth counting honestly, because expectations usually run high.
An agent saves no time on tasks you do once a month. Setting it up will take longer than the task does in a year.
The saving appears with frequency. A ten-minute task repeated daily is around forty hours a year. That is where automation pays back quickly.
And a second, less obvious saving — on switching. A small regular task costs not only its ten minutes: it pulls you out of your work, and getting back in takes as long again. Removing routine gains you more than the arithmetic shows.
What the saving does not give: fewer people. The time freed up goes to what nobody got round to before — and that is a normal and correct outcome.
Where to start
The "AI agents: the electrification of routine" express course — from a first automatic agent to flows working round the clock. Inside: how an agent differs from a chat, twenty ready projects for real tasks, and four working platforms including visual builders and flexible environments.
The price is $9. The course is included in the Make and Team plans — note, that is not the base level, and it is not in the three free days of Basic.
Three routes to it: a one-off purchase at $9, the relevant plan, or 1 000 experience points earned by working on the platform.
What is available sooner and cheaper. Before building agents, it is worth looking at the automation templates — more than 3 000 ready flows. A substantial share of the tasks people take up agents for is already built. The descriptions are open to everyone, the ordinary templates come with Plus, advanced ones with Full.
And 140+ AI assistants across 14 categories, available from Basic: if the task sounds like "I need an assistant on this subject", you probably do not need an agent — you need a ready assistant.
Registration is free and opens three days of Basic — enough to work out whether your task gets solved by something ready before you build your own.
Do one thing today: write out three tasks that repeat every week and require no judgement from you. Those are candidates for automation. The ones needing judgement are candidates for an agent. Separating those two lists is the first practical step.
---
How an agent gets designed in practice
The four-part theory above is what an agent is. Next, how one gets built.
The bot architect from the automations archive designs system prompts: the role, step-by-step reasoning logic along the chain Understand → Basics → Break Down → Analyze → Build → Edge Cases → Final Answer, and a block of hard prohibitions so the agent does not drift into invention.
A prompt generator for multi-agent teams — for when there are several agents. It distinguishes three levels: a Director distributes and consolidates, a Supervisor runs an area, a Basic performs an operation.
Module 9, lesson 1 — the theory of system instructions per section 5.7 of Burkov's book: the key components and why each is needed.
The AI agents mini-course comes with the Make and N8N plans.
The archive — Make, N8N and Full; module 9 — Creator.
---
What to read next
[AI agents: the full breakdown](/en/blog/ai-agents-guide) — how they work, the economics and what people pay.
[Anatomy of a system prompt](/en/blog/system-prompt-anatomy) — how an agent's behaviour gets set.
[Automation with no developer](/en/blog/automation-without-developer) — when rules are enough instead of an agent.
[What an assistant for a department costs](/en/blog/assistant-for-department-cost) — what this turns into as a service.