AI Steps

QuickFlo includes two AI step types that bring large language models directly into your workflows. Use them to generate text, extract structured data, classify content, summarize documents, or run autonomous agents that call other workflows as tools.

Both steps support multiple AI providers — Anthropic (Claude), OpenAI (GPT), and Google (Gemini) — so you can choose the model that best fits your use case and swap between them without changing your workflow logic.

LLM Call

The LLM Call step (ai.llm-call) sends a prompt to an AI model and returns the response. It supports two output modes: text generation for freeform responses, and structured output for extracting typed JSON that matches a schema you define.

Configuration

AI Provider

Field	Description
Provider	Anthropic, OpenAI, or Google
Connection	An API key connection for the selected provider
Model	Select from the dropdown or type a specific model version ID
Enable Web Search	Allow the model to search the web for current information (Anthropic and Google only)

Tip

Models are automatically fetched from the selected provider if you have provided a valid connection. Because of the design time fetching behavior, newly released models for each provider automatically become discoverable in the QuickFlo UI.

Prompt

Field	Description
System Prompt	Sets the AI’s behavior and context — define a persona, rules, or formatting instructions (optional)
Prompt	The user message sent to the model — describe what you want generated or extracted. Supports template expressions.

Knowledge Bases

Select one or more knowledge bases to provide document context. When selected, QuickFlo searches the knowledge bases using the prompt text and injects the most relevant chunks into the model’s context — giving the AI access to your specific documents (RAG).

Attachments

Add images or documents for the model to analyze. Each attachment can be a URL, base64-encoded content, or a data URI. Supported formats include PNG, JPEG, GIF, WebP, PDF, and more.

Output Modes

Text Generation

The default mode — the model generates a freeform text response. Use this for summarization, content generation, classification, question answering, or any task where you want natural language output.

Structured Output

Generate typed JSON matching a schema you define. The model is constrained to return data in exactly the structure you specify — no parsing or cleanup needed.

Define your schema using name-value pairs in the Output Schema section. Each field has a name (the JSON key path) and a value that specifies the type and description:

Value Format	Type	Example
description	string (default)	Business name
number: description	number	number: Phone number
boolean: description	boolean	boolean: Is verified
string[]: description	array of strings	string[]: List of tags
number[]: description	array of numbers	number[]: Monthly revenue figures

Use dot notation and array notation in field names to define nested structures:

Name	Value	Resulting JSON Path
leads[0].company_name	string: Business name	leads[].company_name
leads[0].phone_number	string: Phone number	leads[].phone_number
summary.total	number: Total count	summary.total

Note

Structured output is ideal for data extraction, classification, and any workflow where downstream steps need to process the AI’s response programmatically. The output is guaranteed valid JSON matching your schema.

Advanced Settings

Field	Default	Description
Temperature	0.7 (text) / 0.3 (structured)	Controls randomness — lower values produce more deterministic output
Max Tokens	4,096	Maximum tokens in the response (up to 32,768)
Timeout	120s	Maximum wait time for the model response (up to 5 minutes)

Step Output

Reference the LLM Call output in later steps:

{{ my-llm-call.text }}                    // generated text (text mode)
{{ my-llm-call.object }}                  // structured JSON (structured output mode)
{{ my-llm-call.object.leads }}            // access nested fields
{{ my-llm-call.usage.totalTokens }}       // total tokens consumed
{{ my-llm-call.usage.promptTokens }}      // input tokens
{{ my-llm-call.usage.completionTokens }}  // output tokens
{{ my-llm-call.model }}                   // model ID used
{{ my-llm-call.provider }}                // provider name
{{ my-llm-call.finishReason }}            // "stop", "length", etc.

Example: Data Extraction

Extract structured business data from unstructured text:

Field	Value
Provider	Anthropic
Model	claude-haiku-4-5-20251001
System Prompt	You are a data extraction assistant. Extract ALL businesses mentioned into the leads array.
Prompt	Extract all business leads from this text: {{ search-for-leads.result }}
Output Schema	leads[0].company_name → Business name, leads[0].phone_number → Phone number

The step returns a clean JSON object with a leads array — ready to pipe into a for-each loop, data store, or API call.

---

AI Agent

The AI Agent step (ai.agent) runs an autonomous agent that works toward a goal you define. Unlike a single LLM Call, an agent can reason in multiple iterations and execute other workflows as tools — making decisions about what to do next based on the results of previous actions.

Configuration

AI Provider

Same provider options as the LLM Call step — Anthropic, OpenAI, or Google with connection and model selection.

Goal and System Prompt

Field	Description
Goal	What the agent should accomplish — be specific about the desired outcome. Supports template expressions.
System Prompt	Define the agent’s persona and behavior guidelines — rules, constraints, output format instructions (optional)

The Goal is the primary instruction — it tells the agent what to achieve. The System Prompt sets how the agent should behave while working toward that goal.

Tools

Tools are the agent’s capabilities — other workflow templates that the agent can call to take actions in the real world. Each tool is a workflow that the agent invokes with specific parameters and receives the workflow’s return values as a result.

Field	Description
Tool Workflow	Select a workflow template — any workflow with the Template toggle enabled in Workflow Settings
Tool Name	The name the agent sees — use snake_case for best LLM compatibility (e.g., send_sms, create_ticket)
Description	Tell the agent what this tool does and when to use it — the agent reads this to decide which tool to call
Parameters	Define the input parameters the agent should pass. If not set, parameters are inferred from the workflow template’s Input Schema.

Tip

Write clear, specific tool descriptions — the agent uses them to decide when to call each tool. A description like “Use this tool to send an SMS after finding a lead” is much better than “Sends SMS”.

Knowledge Bases

Attach knowledge bases for the agent to search during its reasoning. The agent can query them autonomously when it needs information to accomplish its goal.

Limits

Field	Default	Description
Max Iterations	10	Maximum number of LLM calls before the agent stops (1–50)
Max Tokens	32,000	Total token budget across all iterations (up to 100,000)
Timeout	120s	Maximum total execution time for the entire agent run (up to 10 minutes)

Advanced Settings

Field	Default	Description
Temperature	0.7	Controls randomness in responses and tool selection
Return Intermediate Steps	Off	When enabled, includes all tool calls and their results in the step output — useful for debugging and auditing

Step Output

{{ my-agent.result }}                  // final response from the agent
{{ my-agent.status }}                  // "completed", "max_iterations", "max_tokens", "timeout", or "error"
{{ my-agent.usage.totalTokens }}       // total tokens across all iterations
{{ my-agent.usage.iterations }}        // number of LLM calls made
{{ my-agent.usage.toolCalls }}         // number of tool calls executed
{{ my-agent.durationMs }}              // total execution time in milliseconds

When Return Intermediate Steps is enabled:

{{ my-agent.steps }}                   // array of all tool calls
{{ my-agent.steps[0].toolName }}       // which tool was called
{{ my-agent.steps[0].arguments }}      // arguments the agent passed
{{ my-agent.steps[0].result }}         // what the tool returned
{{ my-agent.steps[0].durationMs }}     // how long the tool call took

How the Agent Loop Works

1. The agent reads the goal, system prompt, and any knowledge base context
2. It decides what to do — either respond directly or call a tool
3. If it calls a tool, the corresponding workflow template executes with the parameters the agent chose
4. The tool’s return values come back to the agent as a result
5. The agent reads the result and decides its next action — call another tool, call the same tool with different parameters, or return a final response
6. This loop continues until the agent achieves the goal, or hits a limit (max iterations, max tokens, or timeout)

Example: Lead Enrichment Agent

An agent that takes a list of leads, cross-references each one against a CRM, and enriches the records with missing data from the web:

Field	Value
Provider	Anthropic
Model	claude-sonnet-4-6
Goal	For each lead in {{ initial.leads }}, look them up in the CRM. If the lead exists, check for missing fields (phone, company size, industry). Search the web to fill in any gaps and update the CRM record.
System Prompt	You are a data enrichment specialist. Always check the CRM first before searching externally. Only update fields that are empty or outdated. Return a summary of what you enriched.
Web Search	Enabled

Tools:

Tool Name	Workflow	Description
lookup_crm_contact	crm-tools.lookup	Look up a contact in the CRM by email or name. Returns the existing record or null if not found.
update_crm_contact	crm-tools.update	Update a CRM contact record with new field values.

The agent iterates through each lead, queries the CRM to see what data already exists, searches the web to fill in gaps, and writes the enriched data back — adapting its approach based on what it finds for each lead.

---

Choosing Between LLM Call and AI Agent

Use Case	Step
Generate text from a prompt	LLM Call (text mode)
Extract structured data from text	LLM Call (structured output)
Classify, summarize, or transform content	LLM Call
Answer questions using your documents (RAG)	LLM Call with knowledge bases
Multi-step reasoning with tool use	AI Agent
Autonomous research and data collection	AI Agent with web search
Orchestrate multiple workflows based on AI decisions	AI Agent with workflow tools
Tasks where the AI needs to adapt its approach	AI Agent

Note

LLM Calls are faster and cheaper — one model call, one response. Use agents when the task genuinely requires multiple steps, tool use, or adaptive reasoning. Start with an LLM Call and upgrade to an agent only if you need the autonomy.