Go to Studio

AI Agent

Autonomous AI that can use tools and data sources

What does this node do?

The AI Agent node creates an autonomous AI assistant that can reason, plan, and use tools to accomplish complex tasks. Unlike the LLM node which just generates text, the Agent can take actions like searching the web, accessing databases, or calling APIs.

Common uses:

  • Research tasks that require multiple steps
  • Data analysis with tool usage
  • Complex reasoning with external data
  • Tasks that need real-time information

Quick setup

Add the AI Agent node

Find it in AI Nodes → AI Agent

Write your instructions

Describe the task and goal clearly

Enable tools or MCP

Select which tools or MCP the agent can use

Add data sources (optional)

Connect knowledge bases or documents

Configuration

Required fields

instructions string required

The task or goal for the agent to accomplish.

Tips for agent instructions:

  • Define the end goal clearly
  • Specify what tools to use when
  • Set constraints and limitations
  • Describe expected output format

Example:

Research the company {{company_name}} and provide:
1. Company overview
2. Recent news (last 3 months)
3. Key competitors
4. Potential pain points

Use the web scraper to gather information.
Present findings in a structured format.

Optional fields

tools array

List of tools the agent can use during execution.

Available tools:

  • Web Scraper - Fetch and extract web content
  • HTTP Request - Call APIs
  • Search - Search the web

The agent decides when and how to use each tool.

datasources array

Knowledge bases or documents the agent can query.

Add context with:

Here is the data source you will be working with: 
"Company Knowledge Base"

AI settings

Same settings as the LLM node:

model string default: gpt-4o

The AI model to use. GPT-4 or Claude recommended for agents.

temperature number default: 0.6

Lower values (0.3-0.5) recommended for more reliable agent behavior.

max_output_tokens number default: 1000

Maximum tokens for the final response.

output_json_schema object

Schema for structured output.

Output

The agent returns its final response after completing the task:

{
  "response": "Research findings...",
  "tools_used": ["web_scraper", "search"],
  "steps_taken": 5,
  "model": "gpt-4o"
}

How agents work

graph TD
    A[Receive Task] --> B[Plan Steps]
    B --> C{Need Tool?}
    C -->|Yes| D[Use Tool]
    D --> E[Process Result]
    E --> C
    C -->|No| F[Generate Response]
  1. Understand the task from your instructions
  2. Plan what steps are needed
  3. Execute using available tools
  4. Iterate until the task is complete
  5. Respond with the final result

Examples

Company research

Instructions:

Research {{company_name}} and create a company profile.

Include:
- Company description and founding year
- Products/services offered
- Recent news or announcements
- Company size and funding (if available)
- Key competitors

Use the web scraper to visit their website and gather information.
Search for recent news articles about them.

Format the response as a structured report.

Tools enabled: Web Scraper, Search

Competitive analysis

Instructions:

Analyze our top 3 competitors for the keyword "{{keyword}}":

1. Find the top 3 ranking pages for this keyword
2. Scrape each page for content analysis
3. Identify common topics and themes
4. Note any gaps our content could fill

Provide actionable recommendations for outranking them.

Tools enabled: Search, Web Scraper

Data enrichment

Instructions:

Enrich this lead record with additional information:

Lead: {{lead_data}}

Find:
- Company website
- Industry
- Employee count estimate
- Recent company news
- Technology stack (if discoverable)

Use the company email domain to find the website.
Return enriched data as JSON.

Agent vs LLM: When to use which

ScenarioUse AgentUse LLM
Simple text generation❌✅
Content summarization❌✅
Research requiring web access✅❌
Multi-step reasoning✅❌
Data extraction from known text❌✅
Tasks needing real-time data✅❌
High-volume processing❌✅

Best practices

Define clear goals

Tell the agent exactly what success looks like:

❌ "Research this company"

✅ "Research this company and provide a 5-section report 
   including: overview, products, news, competitors, 
   and recommendations. Each section should be 50-100 words."

Limit tool usage

Only enable tools the agent needs:

❌ Enable all tools "just in case"

✅ Enable only Web Scraper and Search for research tasks

Set constraints

Prevent runaway execution:

"Complete this task in no more than 10 tool calls.
If you cannot find the information, say so clearly."

Provide context

Give the agent background information:

"You are researching for a B2B SaaS company selling 
marketing automation tools. Focus on relevant information
for a sales team preparing for outreach."

Common issues

Agent takes too long
  • Reduce scope of the task
  • Set explicit step limits
  • Be more specific about what to find
Agent uses wrong tools
  • Specify which tool to use for what
  • Disable unnecessary tools
  • Add examples to instructions
Results are incomplete
  • Break into smaller sub-tasks
  • Increase max_output_tokens
  • Add “ensure you cover all points” to instructions
Agent loops or gets stuck
  • Add “if stuck, provide best available answer”
  • Set maximum iteration limit
  • Simplify the task
LLM

For simpler text generation
Web Scraper

Extract web content directly