Tool selection

A Tools Agent can only do what its connected tools allow. You choose the tools; the agent chooses which one to call, when, and with what input, based on your prompt and on each tool’s name and description.

How the agent chooses

flowchart TD
  Prompt["Prompt + system message"] --> Read["Model reads the task and the tool list"]
  Read --> Pick{"Is a tool needed?"}
  Pick -->|Yes| Call["Calls one tool with an input"]
  Call --> Result["Reads the result"]
  Result --> Read
  Pick -->|No| Answer["Answers"]

  class Read awf-ai

Each tool call counts toward Max iterations (10 by default). When the limit is reached, the agent answers with what it has.

The model must support tool calling. Most cloud chat models do; for on-device use, see Transformers Chat.

Available tools

Tools that only read or compute

These run without asking you:

Tool	What the agent can do with it
Web Search	Search DuckDuckGo, Bing or Brave and read the first page of results. Needs the browser extension.
Wikipedia	Look up a topic on Wikipedia.
Calculator	Do exact arithmetic instead of guessing.
Think	Write down its reasoning before acting. Helps on multi-step tasks.
Workflow Tool	Run one of your workflows and use its output.

Tools that take actions

These ask for your approval every time the agent calls them:

Tool	What the agent can do with it
Browser	Click, fill, type, select, submit or navigate on the active page.
HTTP Request	Send requests to a web service.
MCP Client	Call the tools of a remote MCP server.

You approve or reject each call in the Executions panel. This can’t be turned off. See Action approval for why.

A connected knowledge base can also work like a tool: set the agent’s Knowledge search option to Let the agent search the knowledge base on its own, and it decides when and what to search.

Wrap your own steps as a tool

The Workflow Tool turns any workflow into a tool. Use it when the agent needs a step no built-in tool provides, such as “get the price from our product page” or “add a row to my sheet”.

Give it a clear Tool Name, such as get_product_price, not workflow_tool.
Write a Tool Description that says what it does, what input it expects, and when to use it.
Keep the workflow small and predictable: the agent decides when to call it, the workflow decides how.

Make choices reliable

Connect only the tools the task needs. Two or three relevant tools work better than ten.
Say in the prompt which tool to prefer. For example: “Use the page text first. Search the web only if it isn’t there.”
Say when to stop. “Stop when you have three sources” is clearer than relying on Max iterations.
Avoid overlapping tools. Two tools that do almost the same thing make the agent hesitate or call both.
Use regular nodes for fixed steps. If a step always happens, such as reading the page before the agent starts, do it with a regular node before the agent, not as a tool.

Example: a price check

flowchart LR
  Trigger["Context Menu trigger"] --> Text["Get All Text"]
  Text --> Agent["Tools Agent"]
  Model["Chat model"] --> Agent
  Search["Web Search"] --> Agent
  Calc["Calculator"] --> Agent
  Agent --> Show["Display Markdown"]

The page text is read by a regular node, because it’s always needed. The agent gets only two tools: Web Search to find other prices, Calculator to compare them.

Ask Aria