9 Automating workflows with agentic AI

 

This chapter covers

  • How language models access and use different types of tools
  • Planning complex tasks and workflows
  • Agent memory and learning over time
  • Frameworks for the implementation of agents
  • Limitations and the future of agents

So far, we’ve mainly learned about the inner workings and applications of predictive and generative AI, which form the foundation of modern AI. Predictive AI analyzes existing data and extracts patterns, while generative AI uses these patterns to produce new data and content. Most of us dream of an AI that automates full workflows and processes, giving us the time and energy to enjoy life and realize our full potential. Our puzzle still lacks some key pieces to manifest this vision. Our AI can’t interact with the external world, learn from these interactions, and strategize and plan for the future. This kind of agentic AI has been on the agenda of research institutions, AI geeks, and tech giants for decades, but it repeatedly runs into severe feasibility limitations.

9.1 Providing language models with access to external tools

9.1.1 Categories of tools

9.1.2 Turning the human into a tool

9.1.3 The ecosystem of tools

9.1.4 Integrating tools with a language model

9.2 Assembling the agent system

9.2.1 The language model as the brain of the agent

9.2.2 Planning the task execution

9.2.3 Learning from memory

9.3 Building at the frontier of AI agents

9.3.1 Common challenges of agent systems

9.3.2 Overcoming the limitations of agent systems

9.4 Trends and opportunities for AI agents

9.4.1 Scaling up with multi-agent collaboration

9.4.2 Chatting with your data

9.4.3 Autonomous enterprise

Summary