Add AI Agent from Actions. Its main input and output connect to the workflow. The four ports underneath supply its capabilities through separate editable nodes. Click a connected node to configure it, or use a plus button to add a missing connection.
| Connected node | What it does |
|---|---|
| Chat Model · required | Chooses OpenAI, Claude or Gemini, Superagento credits or your own API key, with token limits managed automatically. Personal keys are verified and encrypted on the server. |
| Session Memory · optional | Stores the last 1–10 turns in PostgreSQL for a conversation key. Off by default; expires after 24 hours of inactivity and is isolated by account, workflow and agent. |
| Agent Tools · required | Enables calculator, current time, input lookup, reference-text search and up to four fixed HTTP GET endpoints. The model chooses when to use enabled tools. |
| Structured Output · optional | Validates a single record or list of items and required field types. Invalid output gets one repair attempt within the remaining limits. |
- Set the System prompt on AI Agent for its role, rules and response style. Add Task instructions and map Input data, for example {{trigger.body}}. Chat Model links directly to these editors.
- Memory is optional. Set Conversation ID, such as {{trigger.body.session_id}}, to remember messages in that conversation. A blank or missing ID runs the agent without reading or saving history. Use a different, trusted ID for each conversation.
- Connect later actions to the main output. Read the answer as {{steps.AGENT_ID.text}} or structured fields as {{steps.AGENT_ID.json.total}}.
- For a list, choose List of items in Structured Output, then add a Loop before the next action. It detects the list automatically. In Google Sheets, Map current item to columns replaces single-record mappings with fields for each item. All steps after a Loop repeat in order. Empty lists skip those steps; the first error stops the run. Completed external actions remain completed.
- Test the agent. Output result shows the answer, tool arguments and results, model calls, token usage and credits. Testing a sub-node runs its connected agent.
Each port accepts one matching node; Agent Tools groups the enabled tools. Disconnecting memory or output removes that behavior. Disconnecting the model or all tools stops execution. Capability nodes do not incur separate step charges; the agent accounts for each completed model call, including calls before a later failure.
Task: Calculate 125 * 4 + 80 using the calculator. Return a JSON object with total.
Tools: Calculator
Structured Output: total · number · required
Next step: {{steps.AGENT_ID.json.total}}Token and run limits are managed automatically in code; there are no token settings to fill in. HTTP tools read fixed endpoints, allow only listed query parameters and block redirects and private destinations. Saved bearer tokens are restricted to their assigned account and API origin. Use downstream workflow actions for writes or sending messages.