Build a multi-agent AI app with Google Cloud

Multi-Agent Applications with Google’s Agent Development Kit (ADK)

Key Concepts:

• Multi-Agent App: An application built using multiple specialized AI agents working together to achieve a complex task.
• Agent Development Kit (ADK): Google Cloud’s framework for building multi-agent applications, simplifying orchestration and communication between agents.
• LLM Agent: An agent powered by a Large Language Model (LLM), capable of natural language processing and generation.
• Root Agent: The primary agent that receives initial user input and orchestrates the other agents.
• Sub-Agents: Specialized agents that focus on specific sub-tasks within the overall application.
• Vertex AI: Google Cloud’s machine learning platform used for training and deploying custom models.
• Cloud Run: A fully managed compute platform for deploying containerized applications on Google Cloud.

---

Introduction to Multi-Agent Applications & Poet One

The discussion centers around the growing interest in multi-agent applications and demonstrates a practical example built using Google’s Agent Development Kit (ADK). Miguel, a Cloud Engineer at Google, showcases “Poet One,” an application where a user collaborates with an LLM to write a poem. The application’s functionality is demonstrated live, with the host, Martin, participating in the poem-writing process. The initial prompt is “multi-agent apps and ADK,” setting the theme for the collaborative poem.

Agent Roles and Workflow in Poet One

Poet One utilizes several specialized agents, each with a distinct role:

• Muse Agent: Responsible for brainstorming themes, concepts, emotions, visual anchors, and metaphors for the poem. It acts as a creative starting point. A key output is proposing “digital harmony” and “synchronized thought” as core emotions.
• Architect Agent: Suggests a structure for the poem, in this case, couplets to reflect the “synchronized harmony” of the core concept.
• Scribe Agent: Drafts the initial poem based on the inputs from the Muse and Architect agents.
• Critic Agent: Reviews the poem draft, identifying areas for improvement, specifically suggesting a stronger verb than “make.”
• Wordsmith Agent: Revises the poem based on the Critic Agent’s feedback, replacing “make” with “marshall” to convey a sense of action and control, and replacing “symphony” with “currents sharp and fleet” to avoid predictability.

The workflow involves a sequential process where each agent builds upon the output of the previous one, culminating in a refined poem. The final poem reads:

“A silent hand conducts the beat To marshall currents sharp and fleet. The threads of code now join and weave. A harmony our minds conceive. One logic flows from every part Directed by a single heart.”

Technical Implementation & Code Structure

Miguel details the code structure, emphasizing its simplicity thanks to the ADK.

• Root Agent (Poet Coordinator): Defined as an LLM agent object, configured with the model version, a descriptive role, a prompt, defined output variables for inter-agent communication, and designated sub-agents. It receives the initial user request.
• Sub-Agents (e.g., Scribe Agent): Similar code structure to the root agent, including model selection, prompts, output variables, and defined roles.
• Prompts: Crucial for defining agent behavior. Prompts include the agent’s role, overall instructions, step-by-step breakdowns, and input/output formatting.

The ADK handles the orchestration of these agents, eliminating the need for manual “glue code.”

Agent Communication & Data Flow

Agents communicate by referencing each other’s output keys within their prompts. For example, the Scribe Agent utilizes the output from the Muse Agent to generate the initial poem draft. This allows for a seamless flow of information and collaborative refinement.

Advantages of Using the ADK

Miguel highlights three key benefits of using the ADK:

1. Simplified Workflow Definition: The ADK allows defining multi-agent workflows in plain English, minimizing the need for extensive coding.
2. Session Data Management: The ADK automatically manages session data, ensuring each agent has access to the necessary information from other agents.
3. Testing & Troubleshooting: The ADK provides a web tool for testing and debugging agents before deployment.

Why Utilize Multiple Agents?

The core argument for using multiple agents is that specialized agents are more intelligent, particularly when dealing with diverse sub-tasks. Miguel uses the example of a cybersecurity incident response application, where separate agents are needed for analysis, remediation, and communication – skills that shouldn’t be mixed within a single LLM. This specialization prevents the AI from “accidentally mixing up” different skillsets.

Deployment & Learning Outcomes

The application can be deployed to Cloud Run with a single command, making it accessible to other users. Miguel summarizes his key learnings:

• The ADK simplifies the creation of multi-agent applications.
• Focusing AI on single tasks yields the best results.
• Deployment to Cloud Run is straightforward.

Conclusion

The demonstration of Poet One effectively illustrates the power and practicality of multi-agent applications built with Google’s ADK. The framework allows developers to leverage the strengths of multiple specialized LLM agents, creating more sophisticated and effective AI solutions. The emphasis on clear agent roles, streamlined communication, and simplified deployment positions the ADK as a valuable tool for building the next generation of AI-powered applications. The final thought emphasizes the collaborative potential of AI and humans, stating that “AI and humans create the best work when they work together.”