Agents connect generative AI models with everyday users through natural-language interactions.
2
A useful agent combines a model for reasoning, tools for taking actions, and orchestration for memory and state.
3
Multi-agent systems divide work between specialized agents, which can make complex production tasks easier to manage.
Summary
Rukma Sen introduces AI agents as an interface between generative AI models and everyday users. She defines an agent as a system that achieves goals by interacting with its environment, then explains its three main parts: a model that reasons and plans, tools that take actions, and orchestration that manages memory and state. She compares deterministic, generative, and hybrid agents, using examples such as calculators, customer-service chatbots, and financial advisors. She then contrasts single-agent designs with multi-agent systems, where dispatcher, subject-matter, and supervisor agents divide the work. The talk closes with an overview of Vertex AI, including access to models, APIs, Model Garden, and Agent Builder. Sen also connects agent development with responsibilities around safety, privacy, misinformation, and clear communication about generated content.
Agents connect model capabilities to ordinary user interactions
Sen argues that people will often interact with generative AI through an agent. Agents provide a bridge between powerful models and users who do not speak model-language or understand how to work directly with AI systems. Developers decide how an agent interacts with people and what limits surround it. Those choices include making the system ethical, helpful, humane, and safe. Sen compares this change with earlier shifts in computing interfaces, from household machines to internet-connected screens and gestures such as swiping, zooming, and scrolling.
Agent builders have responsibilities around safety, privacy, and generated information
Sen connects the ability to shape AI interactions with responsibility for protecting users. She names ethical questions about moral obligations and guard rails, including preventing misinformation and making clear when an agent has generated content rather than stating an established fact. She also calls out cybersecurity and data privacy. Developers need to consider where data is stored and how information used for reasoning is protected. She says these concerns matter even more as wearables and other forms of unobtrusive computing become more common.
An agent combines a reasoning model, action tools, and orchestration
Sen defines an AI agent as a system designed to achieve specific goals by interacting with its environment. The model is the agent's brain or executive center. It reasons about the goal, determines a plan, and guides behavior. Tools give the agent the ability to act, such as fetching internet data, calling external APIs, booking flights, or processing payments. Orchestration holds the system together by maintaining memory and state and tracking goals. These components let an agent function with some autonomy instead of only generating or reasoning about an answer.
Deterministic, generative, and hybrid agents fit different tasks
Deterministic agents follow fixed rules or algorithms, so the same input consistently produces the same output. Sen uses a calculator as a simple example: 2 plus 2 returns four. Generative agents produce a wider range of outputs and suit tasks that require combining rules or responding creatively. A customer-service chatbot can generate an answer about a product from available company data. Hybrid agents combine both approaches. Sen describes a financial advisor that uses deterministic logic to analyze the market and identify investments, then uses a generative agent to explain the strategy to customers.
A single agent is simple to build but can be unreliable
In a single-agent architecture, one model handles reasoning, planning, and acting. The developer provides instructions and tools for achieving a goal. Sen points to a simple failure mode: when asked to count the letter A in 'banana', a model may give one answer, then change it when asked to check. That behavior can be a serious problem in a production application. The example shows why a straightforward design may need more structure when the system must produce dependable results.
Multi-agent systems divide complex work among specialized roles
Sen compares multi-agent architecture with a company, where people with different roles work toward a shared goal. Her customer-service example has three levels. A dispatcher triages incoming requests, assesses the issue, and routes it. Subject-matter agents handle cases in specific product areas or regions. A supervisor agent checks the work against a predefined data set and can solve some issues itself. Sen also describes a supervisor that rejected every response from the first agent because its instruction to return 'This is not good enough, please try again' created a loop.
Vertex AI offers several ways to build and deploy agents
Sen presents Vertex AI as a full life-cycle AI development platform. Developers can use low-code or no-code approaches as well as highly customized code. Vertex AI provides access to more than 150 models, including Google Cloud models, Anthropic models, Llama 2, Llama 3, and open-source models. APIs support prototyping and experimentation. Model Garden lets users select and fine-tune models and run model evaluation workflows for a specific use case. Agent Builder provides no-code through full-code options for building agents. Sen also describes Google Cloud security, data privacy, compute, and orchestration support for deployment.
"An AI agent simply is a system that's designed to achieve specific goals by interacting with its environment."06:41
Who should watch
You are deciding whether a single-agent or multi-agent design fits a production workflow.
You need a practical definition of agents and want to understand the roles of models, tools, and orchestration.
You are evaluating Vertex AI, Model Garden, or Agent Builder for prototyping and deploying agents on Google Cloud.
