AI Agents Integration in FlowMattic

FlowMattic’s AI Agent is a powerful automation component that integrates multiple AI providers with your workflows. It allows AI models to make decisions, execute actions through tool calling, maintain conversation memory, and return structured responses — all within your automated workflows.

How AI Agents Work

The AI Agent operates in an iterative loop:

1. You provide a system prompt (instructions) and a user message (the task or question).
2. The AI processes the input along with any available tools (FlowMattic actions the AI can execute).
3. If the AI decides to use a tool, it calls it with the required parameters.
4. The tool executes and the result is sent back to the AI.
5. The AI can call more tools or provide a final response.
6. This loop repeats for up to 10 iterations until the AI produces a final answer.

Supported AI Providers

FlowMattic supports 10 AI providers out of the box:

Setting Up a Connection

Each AI provider requires an API key configured in FlowMattic’s connection system.

Getting API Keys

Creating a Connection

1. Go to FlowMattic > Connects in your WordPress dashboard.
2. Click Add New Connection.
3. Search for the AI provider name (e.g., “OpenAI”).
4. Enter your API Key (or Bearer Token, depending on the provider).
5. Click Save to store the connection.

Tip: You can create multiple connections for the same provider — useful if you have separate API keys for different projects or billing accounts.

Configuring an AI Agent

The AI Agent component has three main configuration sections: Chat Model, Memory, and Tools.

Chat Model Configuration

This is the core section where you define the AI model and its behavior.

Required Settings

Model Parameters (Optional)

Fine-tune the AI’s response behavior with these parameters:

Tip: For most use cases, the default parameters work well. Adjust Temperature first — use 0.1–0.3 for factual tasks and 0.7–1.0 for creative tasks.

Output Format

Choose how the AI should format its response:

• Text (default) — The AI returns a plain text response.
• JSON — The AI returns structured JSON matching a schema you define.

When using JSON output format, provide a JSON Schema that describes the expected structure:

{
  "type": "object",
  "properties": {
    "title": { "type": "string" },
    "summary": { "type": "string" },
    "sentiment": { "enum": ["positive", "negative", "neutral"] },
    "tags": {
      "type": "array",
      "items": { "type": "string" }
    }
  },
  "required": ["title", "summary", "sentiment"]
}

OpenAI Responses API

For OpenAI reasoning models like o3-pro, codex-mini, or gpt-5.1-codex-mini, FlowMattic supports the Responses API:

Note: In most cases, leave this set to “No (Auto-detect)”. FlowMattic will automatically switch to the Responses API when needed.

Memory Configuration

Memory allows the AI Agent to remember previous conversations and maintain context across multiple workflow executions.

Memory Types

Memory Settings

Window Buffer Memory:

FlowMattic Tables:

WordPress Options:

Session Identifier

The Session ID determines which conversation thread the AI continues. It works the same across all memory types:

• Leave empty — Auto-generates a unique session ID per execution.
• Pass a specific ID — Continue an existing conversation. Use dynamic variables like {Webhook1.session_id} or {Trigger1.user_id} to group conversations by user or thread.

Example: A chatbot workflow using a webhook trigger can pass {Webhook1.session_id} to maintain separate conversations for each user.

Tools Configuration

Tools allow the AI to execute FlowMattic actions during its reasoning process. This is what makes the AI Agent truly powerful — it can take actions, not just generate text.

Tool Source: FlowMattic Actions

Use any action from any installed FlowMattic app as a tool:

1. Click Add Tool.
2. Select the Application (e.g., HTTP, Email, Slack, WooCommerce).
3. Select the Action (e.g., “Send Email”, “Make HTTP Request”, “Create Post”).
4. Write a clear Tool Description — this tells the AI when and how to use the tool.
5. Configure the action’s parameters (some can be left for the AI to fill dynamically).

Important: Tool descriptions are critical. A well-written description helps the AI understand when to use each tool. For example: “Use this tool to send an email notification to the customer. Requires: recipient email address, subject line, and message body.”

Tool Source: Custom Tools (JSON Schema)

Define custom tools with JSON schemas for integrations not available as FlowMattic apps:

Apps Excluded from Tool Use

The following apps cannot be used as tools (to prevent recursion and control flow issues):

• Control flow: Branch, Router, Filter, Iterator
• Workflow control: Delay, Sub-Workflow
• Triggers: Schedule, Webhook, Webhook Response, MCP Trigger
• Special: Human-in-the-Loop, AI Agent

How Tool Calling Works

When the AI Agent has tools available, it follows this process:

Iteration 1:
  AI receives → System prompt + User message + Available tools
  AI decides  → Call a tool OR respond directly
  If tool called → Execute tool, send result back to AI

Iteration 2-10:
  AI receives → Previous context + Tool results
  AI decides  → Call another tool OR provide final response
  Loop continues until done or max iterations (10) reached

Example: Multi-Step Automation

Scenario: “Research a topic and send a summary email”

1. Iteration 1: AI calls the “HTTP Request” tool to fetch data from an API.
2. Iteration 2: AI receives the API response, processes it, and calls the “Send Email” tool with a formatted summary.
3. Iteration 3: AI receives confirmation that the email was sent and provides a final response: “I’ve researched the topic and sent a summary to [email protected].”

Perplexity — Prompt-Based Tool Calling

Perplexity’s sonar models do not support native function calling. Instead, FlowMattic uses prompt-based tool calling:

• Tool definitions are embedded in the system prompt.
• When the AI wants to call a tool, it outputs: TOOL_CALL: {"name": "tool_name", "arguments": {...}}.
• FlowMattic parses this output, executes the tool, and sends the result back to the AI.

This happens transparently — you configure tools the same way as with other providers.

Using AI Agent in a Workflow

Step 1: Add the AI Agent Step

In your workflow editor, add a new step and select AI Agent as the application.

Step 2: Configure the Agent

Set up the Chat Model, Memory, and Tools sections as described above.

Step 3: Use Dynamic Values

Reference data from triggers and previous steps using FlowMattic’s variable syntax:

System Prompt:

You are a customer support agent for {{company_name}}.
The customer's name is {Trigger1.customer_name} and their email is {Trigger1.customer_email}.
Help them resolve their issue professionally.

User Input:

{Trigger1.support_question}

Step 4: Access Results in Subsequent Steps

After the AI Agent step executes, you can use its output in later steps:

Real-World Examples

Customer Support Chatbot

• Provider: OpenAI (gpt-4o)
• Memory: FlowMattic Tables (persistent, multi-user)
• Tools: Send Email, Create Support Ticket, Look Up Order
• System Prompt: “You are a helpful customer support agent. Always be polite and try to resolve issues. If you can’t resolve it, create a support ticket.”
• Flow: Webhook Trigger → AI Agent → Response sent back via webhook

Content Generation with Structured Output

• Provider: Anthropic (Claude Sonnet)
• Memory: None (one-shot generation)
• Tools: Save to Database, Send Slack Notification
• Output Format: JSON with schema for title, body, tags, and SEO metadata
• Flow: Manual Trigger → AI Agent generates content as JSON → Save to WordPress → Notify team on Slack

Data Processing Pipeline

• Provider: Groq (Llama 4 Maverick) — for speed
• Memory: Window Buffer (short-term)
• Tools: HTTP Request, FlowMattic Tables Insert
• System Prompt: “You are a data analyst. Fetch data from the provided API, analyze it, and store the results.”
• Flow: Schedule Trigger → AI Agent fetches and processes data → Results stored in database

Troubleshooting

Common Issues

“Failed to get API key from connection”

• Verify the connection is active in FlowMattic > Connects.
• Check that the API key is valid and has sufficient credits/quota.

AI not calling tools

• Ensure tool descriptions are clear and specific.
• Check that the system prompt instructs the AI to use tools when appropriate.
• Some models are better at tool calling than others — GPT-4o and Claude Sonnet are recommended.

JSON output is invalid or wrapped in markdown

• FlowMattic automatically strips markdown fences (```json ... ```) from responses.
• If you still get invalid JSON, try a more capable model or simplify your JSON schema.

Tool execution fails

• The AI receives the error and can attempt a different approach.
• Check that the tool’s action is configured correctly (correct connection, required fields).
• Review the workflow execution log for details.

Memory not working

• Verify the Session ID is consistent across executions. If auto-generated, each execution creates a new session.
• For FlowMattic Tables memory, ensure the table exists with the required columns.

Model not found or “not a chat model” error

• For OpenAI reasoning models (o3-pro, codex-mini), enable the Responses API toggle.
• Click the refresh button to re-fetch the latest models from the provider.

Debug Options

Enable these options for troubleshooting:

• Include Execution Log — Logs each iteration step, showing what the AI decided and which tools were called.
• Include Request Body — Records the full API requests sent to the provider.

These are saved in the workflow execution history and can be reviewed from the FlowMattic > Task History page.

Summary