此内容尚不支持你的语言。
An AI agent gets things done through its tools. Besides the dedicated tools (HTTP Request Tool, Calculator, Web Search…), many ordinary steps can be given to an agent as a tool: it then decides when to call them, and with what.
Which steps can be tools
Section titled “Which steps can be tools”| Kind | Steps |
|---|---|
| Apps | Slack, Gmail, Google Sheets, Notion, Linear, GitHub, Todoist, Data Store, Telegram, Google Calendar |
| Browser steps (extension only) | Get All Text, Get Structured Data, Get Selected Text, Screenshot Page, Form Fill, Act on the page, Tab Control, Delete Elements |
In the step picker they appear under Tools · Apps when you add a tool to an agent. The Browser block gives an agent a whole browser in one go.
Add one to an agent
Section titled “Add one to an agent”- Open the agent (a Tools Agent): its Building blocks are at the top of Setup.
- Press + Add a tool and pick, for example, Slack.
- Set it up like the normal step: the connection, the operation (Send message), the fields.
- Give it a Tool name and one sentence for When to use it: the model reads it to decide when to call the tool.
- For each field the model should fill, choose Let the model decide in the field’s label row. The field becomes a dashed The model decides box; write in it what the model should put there (“The channel the user asked for, like #sales”).
Fields you fill yourself stay fixed: the model can’t change them. On each call, the model’s values are put into the “model decides” fields and the step runs as usual, with the same connection, limits and Stop.
Under the fields, The model sees shows the one call the model is offered, for example post_to_slack(channel: text, text: text): the tool name and every “model decides” field with its type (optional ones marked ?), any conflict between them, and when the call asks you first.
In the expression view a “model decides” field shows as {{ $fromAI(...) }}. It only works in a step used as a tool: the workflow doctor flags it anywhere else.
When the agent must ask you
Section titled “When the agent must ask you”Each tool step has Ask before a call:
| Choice | The agent asks… |
|---|---|
| Every call | before every call. Saved and installed workflows without this setting behave this way. |
| Only calls that change things (default for a tool you add by hand) | only for calls that write. Reads (GET, list, search) run by themselves. |
| Skip trusted sites | except for calls to the Trusted sites listed below. Ticking Always allow on this site on an approval card adds a site. |
| Ask me anywhere, even when I’m away | and the run waits for you, up to 24 hours, even in a scheduled or background run. Answer from Runs or a notification. |
Whatever the choice, calls that delete or send as you (send, post, create, reply, publish, invite…) always ask.
When a call needs an answer and nobody is there (a scheduled run), the run waits instead of being refused: it shows in Runs as waiting. Approving continues the agent where it stopped; rejecting tells the model the call was declined. Ask before a call and Trusted sites are cleared when you publish the workflow: whoever installs it decides for themselves.
Safety
Section titled “Safety”- Results that come from outside (web pages, API answers, files) reach the model fenced as data, so instructions hidden in them are not followed.
- After the agent has read untrusted content in a turn, it asks before navigating to a new site or sending a long or random-looking value.
- A derived tool inherits what the step may touch (its sites and connections), shown in the workflow’s touches panel.
See Action approval and Tool selection.