Agent Washing
Agent became a selling word, so products of every kind now use it. That isn't always wrong: a fixed workflow that calls a model can be the right tool. The problem is when the label hides what the product actually does, and buyers expect autonomy, judgment or savings it can't deliver. Anthropic's Building effective agents offers a useful test: in a workflow, models and tools follow predefined code paths, while in an agent the model directs its own process and tool use.
Ask to see the product work on your own data, and ask concrete questions. Which actions can it take, through which tools? Does it decide its own steps, or follow a script? What does it need approval for, and is there a log of what it did? How does it handle a case it wasn't designed for? How was it evaluated, and on what (evals)? A product that's honestly an agentic workflow or an AI copilot can still be worth buying, as long as you know which one you're paying for.
Example
Questions to send a vendor before a demo:
1. Show one task end to end on our sample data. Which steps did the model decide, and which were scripted?
2. List every action the product can take in our systems, and the permissions each needs.
3. Which actions require a person's approval? Can we change that?
4. Where's the log of what it did and why?
5. Show a case where it failed, and what happened next.When to use it
- Evaluating AI products before you buy
- Writing or reviewing your own product's marketing
- Explaining to a team why a tool doesn't behave the way the brochure suggested
When not to use it
- Dismissing a useful scripted automation just because it isn't a true agent

