What is an AI agent? Where the category line falls
An AI agent is a tool that decides its own next step and then takes it, using other software, rather than returning text for you to act on. That single test, does it act without being told each move, is what puts a product in the agents category here. On 28 July 2026 that category held 20 listings.
By the The AI Tools Index team
Agent is the least disciplined word in AI product marketing, and a catalog has to draw the line somewhere or the category swallows everything. The line used here is behavioural rather than architectural: if the product chooses its own next action and executes it against other software, it is an agent. If it produces output for a human to act on, it is not, however capable it is.
What does the test exclude?
That last exclusion is contested, because the strongest coding products now run long chains of edits, tests, and retries without asking. The rule here is where the product's centre of gravity sits: built around a developer's editor and workflow, it is filed under coding, whatever it does internally.
- A chatbot that answers well but never acts. That is a chatbot, and it belongs in that category.
- An assistant that drafts an email for you to send. Drafting is not sending.
- A workflow tool with fixed steps you configured. Automation with a model inside it is still automation.
- A coding assistant that suggests a change you accept. Suggestion is not execution.
Why does the boundary keep moving?
Because the products keep changing shape faster than the vocabulary does. Anthropic's Claude Cowork reached web and mobile on 7 July 2026 after a desktop debut in January 2026, and Google's Antigravity became a standalone application rather than only a development environment. Both changes alter which category a catalog should file them under, and neither was announced as a category change.
The open-source side moves the same way. The project now called OpenClaw previously carried two other names, Clawdbot and Moltbot. A catalog that files by name rather than by identity ends up with the same software listed three times.
What should you actually ask before adopting one?
- What can it do without asking, and what requires approval? Get the list, not the philosophy.
- What credentials does it hold, and can they be scoped down?
- Is there an audit log of actions taken, readable by somebody who is not the operator?
- How does it fail: does it stop, retry, or improvise?
- Can a run be rolled back, and has anyone tested that?
How big is the category, and is that growing?
Twenty listings on 28 July 2026, joint fourth largest in this index alongside coding, behind marketing at 27, productivity at 22, and analytics at 21. More relevant than the size is the churn: agents is where entries are most often renamed, merged into a larger product, or refiled into another category between re-checks. A count in this category ages faster than a count anywhere else here.
The short version
If a product would still be useful with the internet disconnected from everything except its own model, it is an assistant. If disconnecting it from your other software breaks the point of it, it is an agent. That distinction predicts the questions you need to ask about permissions, logging, and rollback far better than any feature list does.