Notes · Note 09
Do you actually need an agent?
Three questions, two minutes. Most things sold as agents should have been one good prompt.
The short version
- One step, no outside data, nothing to resume? It is a prompt, not an agent.
- Agents earn their cost on jobs with many steps that touch other systems and must survive failing halfway.
- Anything running unattended needs hard limits before it runs at all.
“Agent” is doing a lot of work in sales conversations right now, and it is the most common way we see a simple job turned into an expensive, fragile one. Before anyone builds you one, three questions.
1. Is the goal one step or many?
Summarise this document is one step. Research fifteen prospects, write each a note, log them, and schedule follow-ups is many. If it is one step, you want a good prompt and maybe a button — not an agent.
2. Does it need to reach outside itself between steps?
If the job has to look something up, write to your CRM, or check a calendar partway through, that is genuine agent territory. If everything it needs is in the text you already pasted, it is not.
3. Could it fail halfway and need to resume?
A job running twenty minutes across fifteen records will fail partway sometimes. If that matters — if you cannot simply run the whole thing again — you need something that tracks where it got to. That is the real cost of an agent, and the real reason to be sure you need one.
Then build it, and bound it tightly: a hard step limit, a spend limit, a written list of what it may touch, and a switch that stops it. Unbounded agents are how people discover what a runaway loop costs.
Why this matters commercially
An agent costs more to build, more to run, and considerably more to debug at eleven at night. A single well-written call costs cents and fails in ways you can see. Most of the value people expect from agents is available from the simpler thing, and the simpler thing ships this week.
When someone quotes you for an agent, ask them these three questions. The answer tells you whether they are solving your problem or selling their favourite architecture.