A research agent using the DeepAgents harness with planning, sub-agent delegation, and web search capabilities.

Overview

This sample demonstrates the DeepAgents harness, which provides advanced agentic patterns beyond simple tool-calling loops. DeepAgents combines:

• Planning through task decomposition
• Subagent spawning for specialized tasks (researcher and critic)
• File system access for memory management
• Tool use with Tavily search

Requirements

• Python 3.11+
• Anthropic API key
• Tavily API key

Installation

uv venv -p 3.11 .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate
uv sync

Set your API keys as environment variables in .env

ANTHROPIC_API_KEY=your_anthropic_api_key
TAVILY_API_KEY=your_tavily_api_key

Usage

uipath run agent '{"messages": [{"type": "human", "content": "Research the history of artificial intelligence"}]}'

The agent will:

1. Break down complex research questions into sub-tasks
2. Delegate research to the specialized researcher subagent
3. Use the critic subagent to review and provide feedback on findings
4. Use web search to gather information
5. Organize findings into a structured response

How It Works

Unlike simple ReAct agents, DeepAgents can:

• Plan multiple steps ahead using a built-in task decomposition tool
• Spawn specialized subagents with isolated context:
  • Researcher: Gathers information using web search
  • Critic: Reviews outputs for quality and completeness
• Maintain state through filesystem access
• Handle complex, multi-step research workflows with iterative refinement