The development and deployment of agent-based applications of generative AI requires a combination of frameworks, platforms, orchestration tools and cloud infrastructures. Here is a selection of the most relevant tools for creating this type of application.

Frameworks and specialised libraries

These tools facilitate agent management, task orchestration and interfacing with LLMs. These are sometimes usable in multiple environments and available in different languages, most often of course in Python but also in Typescript/Javascript.

LangChain

• Description: Open Source framework project for building reasoning and processing chains with LLM models.
   
• Objective: Allows the creation of agents capable of using tools, interacting with databases and planning tasks.
• Key features:
  • Management of prompts and processing chains.
  • Contextual memory.
  • Integration with APIs and databases.
  • Support for external tools and multi-agent executions.
• Website: https://www.langchain.com/

LlamaIndex (formerly GPT Index)

• Description: Open Source Framework project optimised for connecting LLMs with external data sources (databases, documents, APIs).
• Objective: Crucial for agents capable of retrieving and processing dynamic information.
• Key features:
  • Advanced indexing for contextual search.
     
  • Integration with vectorisation engines (Pinecone, Weaviate, etc.).
  • Construction of optimised queries for RAG (Retrieval-Augmented Generation).
• Website: https://www.llamaindex.ai/

AutoGen (Microsoft)

• Description: Open-source framework designed to manage several agents interacting with each other.
   
• Purpose: Ideal for applications where multiple agents collaborate to solve a problem.
• Key features:
  • Creation of specialised agents.
  • Communication between agents with defined logic.
  • Optimised for GPT-4, Claude, and other LLM models.
• GitHub repository: https://github.com/microsoft/autogen

CrewAI

• Description: Tool based on LangChain for coordinating several AI agents with distinct roles.
• Objective: Simplifies the creation of agents with specific missions, capable of cooperating to achieve an overall objective.
• Key features:
  • Definition of roles and responsibilities for each agent.
  • Management of multi-agent AI teams.
  • Execution of collaborative actions and task planning.
     

• GitHub repository: https://github.com/joaomdmoura/crewAI

   

Vectorisation tools and knowledge storage

AI agents need to store and retrieve information efficiently. These tools enable the creation of vector databases for storing and searching knowledge.

Pinecone

• Description: Vector database enabling efficient storage of embeddings for semantic search.
   
• Purpose: Essential for agents using RAG.
• Alternative: Weaviate, ChromaDB, FAISS.
• Website: https://www.pinecone.io/

Weaviate

• Description: Open-source vector search engine with GraphQL support and integrated ML models.
• Purpose: Useful for AI applications requiring fast retrieval of information.
   
• Website: https://weaviate.io/

Faiss

• Description: Open-source library developed by Facebook AI for similarity search and vector indexing.
• Objective: Optimised to manage millions of embeddings and perform ultra-fast searches on large data sets.
• Key features:
  • High-performance indexing with GPU support
  • Fast approximate search (ANN)
     
  • Optimised for large vector databases
• GitHub repository: https://github.com/facebookresearch/faiss

ChromaDB

• Description: Open-source vector database focused on simplicity and native integration with LLMs.
• Objective: To provide persistent and efficient storage for embeddings with an intuitive API.
• Key features:
  • Local and remote storage of embeddings
  • Optimised search for agents using RAG
  • Integration with LangChain and other frameworks
• Website: https://www.trychroma.com/

Generative AI models and APIs

Agents use LLMs to understand and generate text.

OpenAI GPT-4 Turbo

• Objective: High-performance for agents that require complex reasoning and advanced text generation capabilities.
   
• Website: https://openai.com/

Claude (Anthropic)

• Purpose: Model optimised for security and alignment, perfect for autonomous agents processing sensitive information.
• Website: https://www.anthropic.com/

Mistral & Mixtral

• Purpose: Powerful open-source models, used for AI applications where transparency and cost are important.
   
• Website: https://mistral.ai

Gemini (Google DeepMind)

• Purpose: Multimodal, designed to process text, images and videos in a single query.
• Advantages: Advanced reasoning and contextual analysis capabilities.
• Website: https://deepmind.google/gemini

Llama 3 (Meta AI)
 

• Objective: High-performance open-source model with optimised architecture for efficiency.
• Advantages: Low operating cost and adaptable to applications requiring total control of the model.
• Repo Hugging Face: https://huggingface.co/meta-llama

Command R+ (Cohere)

• Objective: Optimised for RAG with an excellent balance between cost and performance.
   
• Advantages: user-friendly API, optimised for precise responses with knowledge bases.
• Website: https://cohere.com/

Infrastructure and orchestration

Agents must be deployed in a robust and scalable manner.

FastAPI

• Objective: Ideal for exposing agents as fast and lightweight REST APIs.
• Alternative: Flask, Django.
• Website: https://fastapi.tiangolo.com/
   

Ray Serve

• Purpose: Allows for the scalable deployment of LLM agents with distributed orchestration.
• Website: https://www.ray.io/

Docker & Kubernetes

• Purpose: Essential for the deployment of AI applications on the cloud with scalability and efficient resource management.
• Docker: https://www.docker.com/
• Kubernetes: https://kubernetes.io/
   

No-Code / Low-Code tools for agentic AI

For non-technical developers, there are tools that allow you to create AI agents without coding.

OpenDevin (Cognition)

• Purpose: A dev-friendly assistant capable of writing, testing and correcting code.
• GitHub Repo: https://github.com/CognitionAI/devin

Flowise

• Objective: Low-code alternative to LangChain for creating agentic workflows via a graphical interface.
• GitHub repository: https://github.com/FlowiseAI/Flowise

Akkio

• Description: No-code platform for training and deploying AI models without programming experience.
• Objective: Simplify AI integration for businesses by automating analytical and predictive processes.
   
• Key features:
  • Drag-and-drop interface for training models
  • Integration with BI and CRM tools
  • Real-time predictions on data sets
• Website: https://www.akkio.com/

Levity

• Description: Low-code tool for creating intelligent workflows with customised AI models.
• Objective: Enables companies to automate repetitive tasks without human intervention.
   
• Key features:
  • Integration with Slack, Google Sheets, and Zapier
  • Simplified training on specific data sets
  • Automation of AI-based decisions
• Website: https://levity.ai/

Conclusion

The development of generative AI agent applications marks a profound transformation in the way intelligent systems are designed, deployed and operated.
 

This paradigm goes beyond simple automation to introduce agents capable of interacting with their environment, making decisions and learning from their experiences. Thanks to frameworks such as LangChain, LlamaIndex and AutoGen, it is now possible to create agents capable of reasoning, retrieving relevant information and performing complex tasks autonomously.
 

The rise of vector databases such as Faiss and ChromaDB has also provided these agents with an efficient memory, which is essential for improving their adaptability and contextual understanding. Furthermore, the choice of the underlying LLM model, whether GPT-4 Turbo, Claude 3, Gemini, or even open-source models such as Llama 3 and Mixtral, plays a key role in the accuracy and relevance of the results obtained.
 

At the same time, the emergence of No-Code / Low-Code tools, such as Akkio and Levity, is paving the way for a wider adoption of these technologies, making AI accessible to companies that do not have advanced technical expertise. This democratisation is an essential factor in accelerating innovation and integrating AI into various business processes.
 

However, the deployment of these agents in production remains a challenge, requiring a robust infrastructure based on Kubernetes, Ray Serve or FastAPI. It is also crucial to take into account the ethical and governance aspects in order to ensure transparency and effective supervision of these autonomous systems.
 

In conclusion, the evolution of generative AI agentic applications is well under way and promises to revolutionise many sectors, from industrial automation to the optimisation of customer services. The tools and platforms presented here offer a powerful toolbox for experimenting with and building these new intelligent agents. It remains to be seen how these technologies will gradually become part of our daily lives and redefine our interactions with artificial intelligence.