AI

AI agents

Build agents with a chat model, memory and tools, and follow every decision they make.

An AI Agent is a node that uses a chat model to answer each item. It can call tools when it needs information or wants to act, and can remember the conversation.

Build an agent

  1. Add an AI Agent node after your trigger (often a Chat Trigger or a WhatsApp Trigger).
  2. Set the prompt, usually {{ $json.chatInput }} or the incoming message, and a system message that describes the agent's job, tone and rules.
  3. Attach a chat model below the agent: OpenAI, Anthropic Claude, Google Gemini, Mistral, Groq, DeepSeek, OpenRouter, xAI Grok, Qwen, Azure OpenAI, Cohere, Vercel AI Gateway, or a local model with Ollama. You use your own API key for the provider.
  4. Optionally attach memory and tools.

An AI agent on the canvas with a chat trigger, an OpenAI chat model, simple memory and two tools attached below it

A support agent: a Chat Trigger starts it, and the model, memory and tools are attached below the agent.

Memory

Memory lets the agent remember earlier messages in a conversation, using a session key (for example the chat session or the customer's phone number).

  • Simple Memory keeps a rolling window of recent messages in Triggerflow's database.
  • Postgres, Redis or MongoDB Chat Memory store the conversation in your own database.

Tools

Tools let the agent look things up and act:

  • HTTP Request Tool: call any API.
  • Call Workflow Tool: run one of your workflows, so any steps you can build become a tool.
  • Knowledge Base Tool: search documents you added to a knowledge base.
  • Calculator: reliable arithmetic.
  • Posify Tool: products, stock and sales from Posify.

Give each tool a clear description: the agent decides when to use it based on that description.

See what the agent did

Every run shows the agent's messages, which tools it called with which inputs, and what came back. Use this to improve the system message with real examples.

Good practice

  • Tell the agent what it must never do (invent prices, promise refunds…).
  • Keep tools narrow: "look up an order by number" works better than "query the database".
  • For actions that matter, add a human approval step.

Previous: Handle errors · Next: Chat and public chat pages