Four families of AI agents, and the one you actually need

The word "agent" has become a convenient shorthand, and that is exactly what makes it misleading: it covers systems whose complexity, cost and risk vary by an order of magnitude. Before choosing a technology, it is worth knowing which one you mean.

1. The conversational assistant

A language model answers a question without acting on your information system. It is the most common form and the least risky: it drafts, summarises, translates, explains. It cannot break anything, because it does nothing but produce text.

When it fits: first-line support, writing assistance, searching internal documentation. Its limit: it knows nothing about your data unless it is given access, and it closes no loop.

2. The tool-using agent

The same model, but given tools: check stock, create a quote, send an email. It decides when to call them and chains those calls until the request is satisfied. This is the most important step up, because the agent stops talking and starts doing.

When it fits: as soon as a task requires reading from or writing to an existing system. What to plan for: every tool is an open door. The agent's rights must be those of the user who invoked it, never broader, and irreversible actions must go through an approval step.

3. The orchestrated workflow

Here the model no longer decides the sequence — your code does. The path is written in advance (extract, verify, compute, notify) and the model only steps in where judgement is actually useful.

When it fits: a stable, repeated business process where you want consistency rather than initiative. Its decisive advantage: it is predictable. You can test it, measure it, and know why it produced a given result — which matters when the process touches invoicing or payroll.

4. The autonomous agent

You give it a goal rather than a procedure: it explores, tries, corrects itself and keeps going until it reaches the objective. This is the most impressive form, and by far the most demanding.

When it fits: a genuinely open-ended task that cannot be specified in advance. Under what conditions: errors must be recoverable. An autonomous agent is justified where work can be reviewed, tested or rolled back — not where its first mistake reaches a customer.

How to decide

Four questions are enough, and all four must be answered yes before moving up a level:

• Complexity — is the task genuinely hard to specify in advance? If it fits in ten fixed steps, an orchestrated workflow will do it better and cheaper.
• Value — does the gain justify the extra cost and latency?
• Viability — is the model actually good at this kind of task, measured on your own cases rather than on a demo?
• Cost of error — can a mistake be seen, and undone?

The rule that saves the most time is counter-intuitive: start at the simplest level that solves the problem. Most of the needs we meet are handled at levels 1 and 2. Autonomy is not a goal, it is a cost you accept when nothing else will do.

The blind spot: permissions

An agent is never more careful than the rights it is given. That is the question to ask before the one about models: on whose behalf does it act? An agent running with administrator access "to keep things simple" will one day do something nobody authorised — and the log will not say who is responsible.

At TW Micronics we always start there: scope, traceability, reversibility. The rest is engineering.