This repository showcases experiments with the CrewAI framework, demonstrating how to build AI agent teams for tasks like startup idea examination and newsletter generation using both API-based and local LLMs. It targets developers and researchers interested in practical applications of multi-agent AI systems and provides insights into the performance of various LLMs with CrewAI.

How It Works

The project utilizes CrewAI to orchestrate a team of AI agents, assigning them specific roles and tasks. It explores different LLMs, including OpenAI's GPT-4, Gemini Pro, and various models run locally via Ollama, to assess their effectiveness in executing complex workflows and their ability to utilize provided tools. The experiments highlight the challenges and successes in getting agents to adhere to task requirements and leverage their tool-use capabilities.

Quick Start & Requirements

• Install: pip install crewai
• Prerequisites: Python 3.8+, Ollama for local models, API keys for OpenAI/Gemini.
• Setup: Requires configuring API keys and potentially downloading local LLM models via Ollama.

Highlighted Details

• Evaluates OpenAI GPT-4, Gemini Pro, and local models (Mistral 7B, OpenChat 3.5, Nous Hermes, OpenHermes 2.5, Starling 7B, Llama 2 variants, Phi-2) with CrewAI.
• Compares agent performance on tasks like startup idea analysis and newsletter generation with web scraping.
• Documents specific LLM behaviors, including tool usage failures and output quality variations.

Maintenance & Community

This is a personal experiment repository, not an official project with dedicated maintenance or community channels.

Licensing & Compatibility

The licensing is not specified in the README. Compatibility for commercial use or closed-source linking is undetermined.

Limitations & Caveats

The experiments are focused on specific use cases and LLM configurations, and the results may not generalize to all scenarios. Several local models struggled with tool usage, producing generic text instead of executing assigned tasks. The project is marked as "WIP" for the email classifier task.