Skip to content

README.md

Latest commit

 

History

History
251 lines (210 loc) · 8.11 KB

README.md

File metadata and controls

251 lines (210 loc) · 8.11 KB

Multi-Agent Task Execution System

This repository implements a multi-agent system that decomposes complex tasks into discrete steps, routing them to specialized agents for execution. The system comprises three main components:

  1. Planner Agent: Orchestrates the workflow by planning task execution and routing subtasks to worker agents.
  2. Researcher Agent: Gathers information, formulas, and reference materials without performing calculations.
  3. Coder Agent: Executes calculations and evaluates formulas with specific values.

Each agent operates as an independent entry point and can be deployed as a separate process, while still being packaged together as part of an Orchestrator Agent Package.

System Architecture

The system utilizes LangGraph to create a directed graph of agents that can communicate and share state.

Planner Agent Graph

---
config:
  flowchart:
    curve: linear
---
graph TD;
	__start__(<p>__start__</p>)
	input(input)
	create_plan(create_plan)
	supervisor([supervisor]):::last
	invoke_agent(invoke_agent)
	__start__ --> input;
	create_plan --> supervisor;
	input --> supervisor;
	invoke_agent --> supervisor;
	classDef default fill:#f2f0ff,line-height:1.2
	classDef first fill-opacity:0
	classDef last fill:#bfb6fc
Loading

Researcher Agent Graph

---
config:
  flowchart:
    curve: linear
---
graph TD;
	__start__([<p>__start__</p>]):::first
	__end__([<p>__end__</p>]):::last
	__start__ --> researcher___start__;
	researcher___end__ --> __end__;
	subgraph researcher
	researcher___start__(<p>__start__</p>)
	researcher_agent(agent)
	researcher_tools(tools)
	researcher___end__(<p>__end__</p>)
	researcher___start__ --> researcher_agent;
	researcher_tools --> researcher_agent;
	researcher_agent -.-> researcher_tools;
	researcher_agent -.-> researcher___end__;
	end
	classDef default fill:#f2f0ff,line-height:1.2
	classDef first fill-opacity:0
	classDef last fill:#bfb6fc
Loading

Coder Agent Graph

---
config:
  flowchart:
    curve: linear
---
graph TD;
	__start__([<p>__start__</p>]):::first
	__end__([<p>__end__</p>]):::last
	__start__ --> coder___start__;
	coder___end__ --> __end__;
	subgraph coder
	coder___start__(<p>__start__</p>)
	coder_agent(agent)
	coder_tools(tools)
	coder___end__(<p>__end__</p>)
	coder___start__ --> coder_agent;
	coder_tools --> coder_agent;
	coder_agent -.-> coder_tools;
	coder_agent -.-> coder___end__;
	end
	classDef default fill:#f2f0ff,line-height:1.2
	classDef first fill-opacity:0
	classDef last fill:#bfb6fc
Loading

Agent Responsibilities

  • Planner Agent:

    • Takes user questions and formulates execution plans.
    • Routes tasks to appropriate worker agents.
    • Manages execution flow and state tracking.
    • Returns final results to the user.

  • Researcher Agent:

    • Retrieves information using a Tavily search tool.
    • Provides factual content, definitions, and formulas.
    • Does not perform calculations (strictly enforced).

  • Coder Agent:

    • Executes calculations using Python code.
    • Evaluates formulas with specific input values.
    • Returns precise numerical results.

Usage

Running the System

To run the entire system, use the following command with the complete JSON input:

uipath run planner '{"question": "First, please state the Pythagorean theorem. Give only the formula, using variables a, b, and c. Then apply this formula to calculate the value when a=2 and b=3."}'

Debugging Individual Agents

You can debug individual agents by invoking them directly:

Researcher Agent

Run the researcher agent with:

uipath run researcher '{"messages":[{"content":"State the Pythagorean theorem formula using variables a, b, and c","type":"human"}]}'

Coder Agent

Run the coder agent with:

uipath run coder '{"messages":[{"content":"Let me help you state the Pythagorean theorem formula. The Pythagorean theorem is a fundamental mathematical formula that describes the relationship between the sides of a right triangle.\n\nThe formula is:\n\na² + b² = c²\n\nWhere:\n- a and b are the lengths of the two legs (the sides adjacent to the right angle)\n- c is the length of the hypotenuse (the longest side, opposite to the right angle).","type":"human","name":"researcher-agent"},{"content":"Calculate the result using the formula when a=2 and b=3","type":"human"}]}'

Sample Workflow

  1. User submits a question about the Pythagorean theorem.
  2. Planner creates an execution plan with two steps:
    • Step 1: Researcher agent retrieves the Pythagorean theorem formula.
    • Step 2: Coder agent applies the formula to calculate the result for a=2, b=3.
  3. Planner executes Step 1 by invoking the researcher agent.
  4. Researcher agent returns the formula a² + b² = c².
  5. Planner executes Step 2 by invoking the coder agent.
  6. Coder agent calculates c = √(2² + 3²) = √(4 + 9) = √13 ≈ 3.606.
  7. Planner combines the responses and returns the final answer to the user.

Steps to Execute Project on UiPath Cloud Platform

  1. Clone the Repository

    git clone https://github.com/UiPath/uipath-langchain-python.git

  2. Navigate to the Sample Directory

    • Windows:

      cd .\uipath-langchain-python\samples\multi-agent-planner-researcher-coder-distributed

    • Unix-like Systems (Linux, macOS):

      cd ./uipath-langchain-python/samples/multi-agent-planner-researcher-coder-distributed

  3. Create and Activate a Virtual Python Environment

    pip install uv
uv venv -p 3.11 .venv
.venv\Scripts\activate  # Windows
source .venv/bin/activate  # Unix-like Systems
uv sync

  4. Authenticate with UiPath Cloud Platform

    uipath auth

    Note: After successful authentication in the browser, select the tenant for publishing the agent package.

    Received log data
Received authentication information
Available tenants:
  0: cosmin
  1: DefaultTenant
  2: Demo
  3: lucian
  4: Solutions
  5: SolutionsTest
  6: Test
  7: TestRoles
Select tenant: 2

  5. Package and Publish Agents

    uipath pack
uipath publish

    Note: You will need to select the feed for publishing:

    Select feed type:
  0: Tenant package feed
  1: Personal workspace
Select feed: 0

  6. Create Agent Processes in Orchestrator

    • Planner Agent planner-agent-package-overview planner-agent-process-configuration

    • Researcher Agent researcher-agent-package-overview researcher-agent-process-configuration

    • Coder Agent coder-agent-package-overview coder-agent-process-configuration

    Note: Ensure that the display names for the coder and researcher agent processes are coder-agent and researcher-agent, and that all 3 agents are placed in the same folder.

  7. Run the Planner Agent with Any Input Question

    Tip: For a five-step action plan, consider using the following input:

    Could you find a Python solution for the N-Queens puzzle for N=8? Please analyze why this solution works,
considering the key programming concepts it employs.
Then, revise the solution to handle a dynamic value of N, where N is any positive integer.
After that, research the time and space complexity of this new solution.
Lastly, demonstrate this revised Python solution with N=10.

Implementation Details

The system is implemented using:

  • LangGraph for agent orchestration
  • LangChain for agent creation and tool integration
  • Claude 3.5 Sonnet for LLM capabilities
  • Tavily for search functionality in the researcher agent
  • Python REPL for code execution in the coder agent

Requirements

  • Python 3.10+
  • LangGraph
  • LangChain
  • Anthropic API key
  • Tavily API key