Agents and automation platforms

What the word means on a product page

“Agent” now labels at least four different things, and knowing which one you are buying prevents most disappointment.

A chat assistant with tools. It can search, run code, read your files. Useful, well understood, and not autonomous in any interesting sense.

A workflow with a model in it. A defined sequence (new email arrives, extract fields, write to a spreadsheet, send a reply) where a model does one or two steps. Reliable, because the sequence is fixed and only the judgement is delegated.

A loop that decides its own steps. Given a goal, it plans, acts, observes and repeats until it thinks it is done. This is what the research community means, and it is the least reliable of the four.

A hosted product with a chat box. Marketing.

Most business value today is in the second category, and most of the excitement is about the third.

Why the autonomous loop is hard

Reliability compounds badly. If each step is right 95% of the time, ten steps in sequence succeed about 60% of the time; twenty steps, about 36%. Errors do not merely accumulate, they mislead: a wrong observation at step four sends the plan somewhere useless for the next six.

This is not an argument against agents. It is an argument for short loops with checkpoints: three or four steps, then a result a person or a rule can verify, then the next three. Every reliable system built this way looks less impressive in a demonstration and works far better on a Tuesday.

The no-code platforms

Zapier and Make connect hundreds of services with a model step available. Fast to build, priced per task run, and the price grows with volume in a way that surprises people at the second invoice.

n8n is the free, self-hostable equivalent. More setup, no per-run cost, and your data stays where you put it. For anything running at volume, or touching material you would rather not route through a third party, this is the one to look at.

Power Automate and the equivalents inside office suites, which are already licensed in many organisations and often overlooked.

Building the same workflow in a scripting language is also free, more flexible, and harder to hand over to a colleague. The platforms earn their place mainly by being maintainable by somebody other than the person who built them.

The questions to ask before switching one on

1. What can it do that cannot be undone? Send mail, post publicly, move money, delete records. Those steps need confirmation, always.
2. What does it do when it fails? Stop and alert, or retry silently forever?
3. Is there a rate limit on its own actions? An agent in a loop can send four hundred emails.
4. Who sees the log? You need to be able to reconstruct what it did and when.
5. What does it read? If it processes incoming mail, web pages or tickets, that content can contain instructions aimed at it. An agent reading attacker-supplied text with permission to act is the central unsolved security problem in this area, and it has no complete fix today: the practical mitigations are least privilege, confirmation on consequential actions, and not letting one agent both read untrusted input and act on sensitive systems.

Cost, which behaves differently here

Automation platforms charge per task run, and a workflow that seemed cheap in testing changes character at volume. A run that costs a fraction of a penny is nothing at fifty a day and a real bill at fifty thousand a month, and the model call inside it is charged separately on top.

Work out the monthly cost at your expected volume before building, then double it, because workflows acquire steps. Self-hosting removes the per-run charge entirely and replaces it with a server to look after, which is the right trade above a few thousand runs a month and the wrong one below it.

The sensible starting point

Automate something small, boring and reversible. Sorting incoming messages into folders. Drafting (not sending) a reply. Extracting fields from attachments into a spreadsheet. Watch it for a fortnight before extending it.

The failure that costs people is the ambitious first project: a system that touches customers, money or public output before anybody has learned how it behaves when it is wrong. It will be wrong, on a schedule you did not choose, and the only question that matters is what it is holding at the time.