Skip to main content

Module 21.5: MAS Knowledge Milestone - Architecture Choice

Theory​

Congratulations! You have completed the core Multi-Agent Systems (MAS) section of the course. You have moved from thinking about single agents to designing complex networks of specialized nodes.

Before we move on to operational concerns like State, Evaluation, and Observability, let's recap the different architectural patterns you've mastered. Choosing the right geometry for your graph is the most important decision an AI Architect makes.

MAS Architecture Recap​

PatternModuleKey ADK PrimitivesBest For...
Static Orchestration16Workflow, Linear Edges, JoinNodePredictable, fixed business processes with parallel steps.
Structured Routing17Workflow, Conditional Edges (Dict), a small @node to set ctx.routeWorkflows that branch based on deterministic model outputs.
Dynamic Orchestration18@node, ctx.run_node()Complex logic, nested loops, and code-driven decision making.
Collaborative Teams19sub_agents, mode="task"Fluid hand-offs where specialists manage their own task lifecycle.
Cyclic Workflows20@node(rerun_on_resume=True) with a Python for/while loop calling ctx.run_node()Iterative refinement, self-correction, and "human-in-the-loop" review loops.
Distributed Graphs21RemoteA2aAgent, to_a2a()Independent scaling, team-separated codebases, and cross-org collaboration.

How to choose?​

When designing a new system, ask yourself these three questions:

  1. Is the path predictable? If yes, use Static/Structured edges for performance and clarity.
  2. Does the logic require Python control flow (loops, try/except)? If yes, use Dynamic (@node) workflows.
  3. Do the agents need to live in different environments? If yes, use Distributed (A2A).

Key Takeaways​

  • There is no "one size fits all" architecture. Most production systems use a Hybrid approach.
  • Always prefer the simplest geometry that solves the problem. Don't use a Dynamic workflow if a Static one suffices.
  • Use the Dev UI Graph View to validate that your code matches your mental model.