Agentic Workflows Project

This project implements a small multi agent workflow that generates a short research report from a topic. It shows how to combine planning, tool based research, writing, and editing in one end to end flow.

What this project does

Given a topic, the system runs a simple pipeline:

1. A planner agent creates a short plan as a list of steps
2. An executor agent reads each step and chooses the right agent to run it
3. A research agent gathers external info using tools (arXiv, web search, Wikipedia)
4. A writer agent drafts parts of the report
5. An editor agent critiques and improves the draft
6. The final output is a Markdown research report

Goal

The goal is to build an agentic workflow that is reliable and easy to extend.

This includes:

• writing prompts for different agent roles
• enabling tool use for research
• orchestrating multiple steps with shared context
• using feedback to improve a draft

Why it matters

Many real ML and data products need more than one single model call. This project is a simple example of how to structure multi step LLM work so it can:

• run repeatably
• keep a history of what happened at each step
• produce a final output that can be reviewed and improved

What is included

• Agentic_Workflows.ipynb
  • planner_agent generates a plan as a Python list
  • research_agent uses tool calls to gather sources
  • writer_agent drafts technical text
  • editor_agent revises and improves drafts
  • executor_agent coordinates the workflow (the orchestrator)

Unit tests are included in the notebook to validate each agent function.

Requirements

You need:

• Python 3.x
• Jupyter Notebook
• aisuite
• the local modules used by the notebook:
  • research_tools
  • unittests

If your research_tools uses external APIs (example Tavily), you may need the related API key in your environment variables.

How to run

1. Open Agentic_Workflows.ipynb
2. Run the cells in order
3. Run the unit test cells to confirm each agent works
4. Run the final example that calls executor_agent(topic)

The executor limits the plan to a maximum of 4 steps by default to keep runtime reasonable.

Output

The final step returns a Markdown research report for the chosen topic, with references gathered by the research tools.