Foundations of multi-agent systems with ADK

Key Concepts

• Multi-Agent System (MAS): Decentralized, local view, emergent behavior.
• Google Agent Development Kit (ADK): LM Agent, Workflow Agent (Sequential, Parallel, Loop), Custom Agent.
• Agent Hierarchy: Parent-subagent relationship, single-parent rule.

1. What is a Multi-Agent System?

• Definition: A system where individual agents work together to achieve a common goal.
• Decentralized Control: No central authority dictates actions; each agent makes independent decisions. Analogy: A flock of birds with no leader, yet forming complex patterns.
• Local View: Agents have limited knowledge of the overall system and react based on their immediate surroundings. Analogy: A person in a crowded stadium only reacting to nearby individuals.
• Emergent Behavior: Complex, global patterns arise from simple, local rules and interactions. Agents collaborate to solve problems beyond the capability of a single agent.

2. How ADK Supports Multi-Agent Systems

• Google Agent Development Kit (ADK): Designed to facilitate multi-agent system development by providing specific agent types.
• LM Agent: The "brain" of the system, using large language models (like Gemini) to understand user input, reason, and decide on the appropriate tool or action.
• Workflow Agent: Orchestrates the interaction and workflow between agents to complete tasks. Three types:
  • Sequential Agent: Executes agents in a linear sequence, like an assembly line.
  • Parallel Agent: Runs multiple tasks concurrently, similar to calling several APIs in parallel.
  • Loop Agent: Repeats a task until a specific condition is met.
• Custom Agent: Allows developers to write custom Python logic by inheriting from a base agent, providing full control over agent behavior.

3. Agent Hierarchy

• Hierarchical Structure: Agents are organized in a hierarchy, similar to an organizational chart.
• Parent-Subagent Relationship: Parent agents manage multiple subagents.
• Single-Parent Rule: Each subagent has only one parent agent.
• Analogy: A company structure with a CEO (root agent), VPs, managers, and workers, each with defined roles.

4. Synthesis/Conclusion

The video introduces multi-agent systems, emphasizing their decentralized nature, local perspectives, and emergent behaviors. It highlights how Google's ADK simplifies the development of such systems by providing LM agents, workflow agents (sequential, parallel, and loop), and custom agents. The concept of agent hierarchy is presented as a foundational element for structuring these systems, enabling efficient collaboration and task completion. The next episode will delve deeper into workflow orchestration and agent communication.