AI Agent node
A workflow node that researches a task with web and news search, Wikipedia and page reading, then writes an answer with sources.
What it does
The AI Agent node answers questions that need fresh information, such as What are the top 5 technology news stories today?. It decides which tools to use, reads what they return, and writes an answer with links to its sources.
- Task
- Agent picks a tool
- Reads the results
- Repeats if needed
- Answer with sources
Use it to feed real, current facts into the rest of a workflow, for example a daily news summary that becomes an Instagram caption and a Telegram message.
Settings
| Setting | What it does |
|---|
| Task | What you want to find out. Supports expressions such as {{ $vars.topic }}. Text wired into Task In is added as context. |
| Extra instructions | Optional. Tone, length or audience, for example Answer in 5 short bullet points. |
| Model | Only models that can call tools are listed. |
| Search results from | Today, this week, this month or any time. |
| Most tool calls | 1 to 10. More calls give deeper research but take longer and cost more. |
| Answer format | Markdown, plain text, bullet list, JSON (answer and sources fields) or Telegram (HTML for the Telegram node). |
Outputs
- Answer carries the formatted answer text. Wire it into a caption, a Telegram message or the AI Text Generator.
- Sources and steps carries everything:
answer, sources (title and link), steps, toolCalls, tokensUsed and partial. - In expressions, use
{{ $node["AI Agent"].answer }} or {{ $node["AI Agent"].sources }}.
Try it
- Add the node: In a workflow, add AI Agent from the Source group and double-click it.
- Write the task: For example
What is the latest news in AI today? Give 5 headlines with one line each. - Run agent now: Click Run agent now. Each tool call appears as it happens, then the answer and its sources.
- Connect it: Wire Answer into a Telegram or Instagram node and schedule the workflow to get it every morning.
Cost and limits
- Each run costs 10 tokens plus 2 per successful tool call. The most it can cost is held when the run starts, and anything unused is refunded when it finishes.
- If a run fails, every token is refunded.
- A run has about 50 seconds. If time or tool calls run out, you get the best answer so far, marked as partial.
- Up to 10 runs a minute, and 3 at a time, per account.
Check facts before you publish. The agent quotes what it finds on the web, and web pages can be wrong. Open the sources for anything important.