What actually makes something an agent
The word gets stuck on everything. Here's a simple test.
Vendors call almost anything an agent now. A chatbot with a nicer name, a fixed automation with one AI step: they all get the label. That makes it hard to buy or build the right thing.
The test
Ask who decides the next step. If a person or a fixed workflow decides, it's an assistant or an automation. If the model chooses which tool to use next based on what it just found, and keeps going until the job is done, it's an agent.
Why you should care
That choice changes your risk. A fixed workflow fails in predictable places. An agent can take a path nobody planned, so it needs tighter permissions, step limits and logs.
When a vendor says "agent", ask: what can it do without approval, what stops it, and can I see what it did? The answers will tell you more than the label.
Three labels, one example
- Assistant: you ask for a draft reply to a customer, and you decide what happens to it.
- Automation: every new ticket is classified by AI, then a fixed rule routes it.
- Agent: given a billing complaint, it decides to check the invoice, then the payments, then the emails, and stops when it has an answer.
Three questions for any agent demo
- Show me a run that failed. What happened?
- What stops it from taking an action it shouldn't?
- Where can I read everything it did?
A serious team will have answers ready.