What It Actually Means
The distinction from a chatbot is doing versus saying. A chatbot answers. An agent looks up the order, checks the stock, issues the refund and emails the customer.
That distinction is the whole story commercially. Answers can be ignored. Actions have consequences, and the interesting question about any agent is not how clever it is but what it is permitted to do without asking.
Where They Genuinely Work
Agents earn their keep on work that is repetitive, rule-shaped, and involves several systems that do not talk to each other.
The realistic candidates look like this: gathering information scattered across tools and assembling it into one place; triaging incoming requests by working out what each is about and routing it; monitoring for a condition and acting when it occurs; handling the first stage of a request end to end and passing on anything unusual.
The common thread is that a person could do it, it takes them a while, it requires judgement too fuzzy for simple rules, and getting it wrong occasionally is survivable.
Where They Fall Over
Long chains. Each step carries a chance of being slightly wrong, and errors compound. An agent taking twelve steps has many opportunities to drift, and it will pursue the drifted goal with the same confidence. Short chains with checkpoints are far more reliable than long autonomous runs.
No sense of consequence. An agent does not distinguish sending an internal note from emailing four hundred customers. Both are just an available action. Anything irreversible needs to be gated by design rather than by instruction, because instructions are guidance and permissions are enforcement.
Edge cases. Agents handle the common path well and the unusual case confidently and wrongly. The unusual case is often the one that mattered.
Silent failure. An agent that quietly does nothing looks identical to one with nothing to do.
What To Ask
- What can it do without asking, and what requires approval? The most important design decision, and it should be a list you have seen.
- Can it spend money, contact customers, or change records? Each deserves a separate deliberate answer.
- What does it do when unsure? Stopping and escalating is the correct behaviour and it has to be built in.
- How would we know it was doing the wrong thing? Logging what it did and why, in something a person can read.
- How do we undo a run? Sometimes trivial, sometimes impossible. Worth knowing which before it runs.
The Honest Framing
Agents are not staff replacements and the pitch that says so is the one to be wary of. What they are is a way to automate work that was previously too fuzzy for rules and too small to justify a person’s full attention.
The failures are rarely the model being incapable. They are scope: an agent given a goal broader than anyone thought through, with permissions wider than the goal needed, and nobody watching what it did.
Start with something narrow, reversible and observed. Widen it once it has earned that. We build these as Beacon Agents, and the first conversation is always about what an agent is allowed to do rather than what it can do.
More terms are in the glossary.
