Skip to main content

Lab 21.5 Solution: Architecting Multi-Agent Systems

Goal​

This solution provides the architectural blueprints and justifications for the scenarios presented in the Milestone Challenge.

Recommended Pattern: Hybrid Static + Structured Routing

  • Design:

    1. START -> Extractor
    2. Extractor -> PrivacyChecker (Edge 1)
    3. Extractor -> LiabilityChecker (Edge 2)
    4. PrivacyChecker -> JoinNode
    5. LiabilityChecker -> JoinNode
    6. JoinNode -> ReviewRouter (Structured Routing via Dictionary)
    7. ReviewRouter -> SeniorPartner (IF high risk)
    8. ReviewRouter -> Summarizer (IF low risk)

  • Justification:

    • Performance: Using parallel edges and a JoinNode ensures that both legal checks happen concurrently, minimizing wait time.
    • Predictability: The logic for "High Risk" vs "Low Risk" should be deterministic (Structured Routing) to ensure every contract follows the exact legal procedure.

Scenario 2: The Multi-Turn Story Writer​

Recommended Pattern: Cyclic Workflow (Module 20)

  • Design:

    1. START -> refinement_orchestrator — a single @node(rerun_on_resume=True) function. The Workflow's edges are simply [("START", refinement_orchestrator)]; there is no edge that loops back to a previous node.
    2. Inside refinement_orchestrator, a plain Python for loop repeatedly calls ctx.run_node(critic, current_story) and, unless the critic responds "APPROVED", calls ctx.run_node(refiner, ...) to produce the next draft.
    3. The loop breaks as soon as the critic returns "APPROVED" (or a max_iterations cap is reached), and the function returns the final story.

  • Justification:

    • Iteration: This scenario requires a feedback loop. Rather than modeling the cycle as a graph edge that returns to a previous node, ADK 2.0 implements it with standard Python control flow (for/while) inside one orchestrator node, calling ctx.run_node() on critic/refiner for each pass -- simpler to reason about and debug than a literal cyclic graph.

Scenario 3: The Global Enterprise Support Bot​

Recommended Pattern: Distributed Graphs (A2A - Module 21)

  • Design:

    1. START -> WebOrchestrator
    2. WebOrchestrator -> RemoteA2aAgent (EU Logistics)

  • Justification:

    • Security & Ownership: Since the EU agent is in a different project and managed by a different team, the A2A protocol is mandatory. It allows the main bot to delegate tasks securely over the network without needing access to the EU agent's source code or private project resources.

Self-Reflection Answers​

  1. Hybrid Approach: Real systems often start with a static business process (Static) but then encounter "messy" data that needs Python logic (Dynamic) or external expertise (Distributed). Hybrid designs offer the best balance of control and flexibility.
  2. Collaborative Team Risks: In a regulated environment, "Collaborative" agents might deviate from a strict protocol or lose the "Chain of Custody." For financial/legal tasks, Static/Structured graphs are preferred for auditability.
  3. Graph Mental Model: Business leaders understand flowcharts and process maps. Explaining an AI system as a "Graph of Nodes" makes the ROI and process logic much clearer than just saying "it's an intelligent chat."