AI Agent Brief by Google

Google released a white paper on AI Agents in September 2024, which I summarized based on the work of Julia Wiesinger, Patrick Marlow, and Vladimir Vuskovic.

GENERATIVE AI MODELS, LIKE HUMANS, CAN BE TRAINED TO ACCESS INFORMATION OR SUGGEST ACTIONS IN REAL TIME, BUT THEY REQUIRE ACCESS TO EXTERNAL TOOLS AND THE ABILITY TO PERFORM TASKS ON THEIR OWN.

• To find specific information, you can extract from the database or use the
• Personalized shopping lists based on purchase history

You can.

Agents combine thinking, logic, and access to external information to expand the scope of generative AI, which used to be standalone.

‍

What is an agent

• LLM MODELS ARE THE BRAINS OF THE AGENT PROCESS
  • Thinking with logic frameworks like ReAct, Chain-of-Thought, and Tree-of-Thoughts
• Tools is where the agent accesses external data and services.
  • FOR EXAMPLE, MAKE YOUR API ACCESSIBLE
• The Orhestration Layer is responsible for periodically controlling the agent's process.
  • A kind of workflow concept
• Difference between agents and models

‍

How agents work

• Gather information
• Reasoning and
• Run an action

ReAct: A Prompted Engineering Framework for Generative Models

Chain of Thought (CoT): A prompted engineering framework to help you think through intermediate steps

Tree-of-thoughts (IoT): A prompted engineering framework for strategic preview or exploration.

Tools

Extension

A role that sits between an agent and an API for executing actions, which may have a Built-in-Extension depending on the type or type of API.

Functions

From a developer perspective, the concept of implementing an extension as a module of code, decoupled from the agent and dependent on the application it actually connects to

Data stores

The ability to utilize externally updated information, often implemented as a vector database. The most well-known model is the Retrieval Augmented Generation (RAG) implementation.

Tools at a glance

‍

‍

How to improve model performance

• In-context learning: define specific conditions under which an agent will run, leveraging the ReAct framework and more.
• Retrieval based in context learning: A technique for leveraging RAGs to improve model performance.
• Fine-tuning based learning: ideal for cases that utilize large datasets tailored to specific cases

Summary

Building complex agent architectures requires an iterative approach, with experimentation and refinement tailored to specific business cases and organizational needs being key. Due to the generative nature of the models that underpin these architectures, no two agents are alike. However, by leveraging the strengths of each component, you can extend the capabilities of the language model and create an agent program that delivers real value

‍

‍