Summary
You started this module facing Contoso Capital's scaling challenge—flat orchestration patterns couldn't coordinate 12 specialized agents with complex dependencies and partial failure requirements. You explored three architectural patterns that solve enterprise orchestration problems: hub-and-spoke for centralized workflow control, hierarchical for domain boundary isolation, and supervisor for quality-gated progression.
The key architectural decision is pattern selection. Use hub-and-spoke when workflows are deterministic with clear spoke invocation sequences—you know that investment research always runs market analysis before compliance checking. Choose hierarchical orchestration when organizational boundaries create natural agent clusters—market analysis specialists operate independently from compliance specialists, and coupling them through one hub creates bottlenecks. Adopt supervisor patterns when agent output quality varies and you need validation gates—external API calls sometimes return stale data that shouldn't propagate downstream without review.
For parallel execution, you learned three synchronization patterns: fan-out with asyncio.gather for full synchronization barriers, quorum-based policies for partial failure tolerance, and semaphore-limited concurrency for quota management. The critical implementation detail is per-agent timeout configuration—divide your total deadline across sequential orchestration rounds, leaving buffer for coordination overhead. Partial failure handling uses three strategies: fail-fast when all agents are critical, best-effort when partial results have value, and minimum-quorum when regulatory or quality thresholds exist.
Framework selection balances ecosystem integration with control flow preferences. Microsoft Agent Framework provides native managed identity support, durable agent persistence, and Microsoft 365 plugin integration—choose this when Azure ecosystem fit matters. LangGraph offers explicit state graph control flow with conditional routing—choose this when regulatory compliance requires auditable execution traces. AutoGen supports collaborative agent reasoning through group conversations—choose this when agents critiquing each other improves output quality. CrewAI provides high-level task delegation abstractions—choose this for rapid prototyping when development speed matters more than fine-grained control.
The underlying insight is framework independence. Hub-and-spoke, hierarchical, and supervisor patterns exist regardless of whether you implement them with Semantic Kernel, LangGraph, AutoGen, or CrewAI. The patterns solve coordination problems; frameworks provide implementation conveniences. Select the pattern based on your workflow requirements and failure tolerance needs, then choose the framework based on ecosystem integration and team skills.
In Contoso Capital's production platform, you implemented hierarchical orchestration with parallel market analysis execution and quorum-based failure policies. When one market data feed times out, the minimum-quorum policy allows research to proceed with three of four data sources rather than failing entirely. The compliance coordinator acts as a quality gate, using supervisor patterns to validate combined analysis before final report synthesis. This architecture handles the enterprise requirements that flat patterns couldn't address: partial failures degrade gracefully, agent clusters evolve independently, and complex dependencies coordinate through explicit hierarchical control.