Why decomposing a task across specialized subagents outperforms one generalist agent, orchestrator vs. peer-to-peer topologies, the LangGraph vs. CrewAI framework landscape, shared vs. private memory, coordination failure modes, and the real cost and latency budget more agents actually cost.
Module 7 covered the single agent; this module starts from the point where one agent, no matter how well built, hits a scope it cannot reliably hold alone. Splitting a complex task across several narrower, specialized subagents is task decomposition — and this module works through the two topologies that coordinate them (a central orchestrator vs. peer-to-peer negotiation), the real framework landscape (LangGraph's explicit graph-based control flow against CrewAI's role-based crew abstraction), and the shared-vs-private memory question that only exists once more than one agent is involved.
By the end of this module you should be able to name the coordination failure modes specific to multi-agent systems — agents talking past each other, redundant work, one agent's error silently propagating into another's context — and explain why every additional subagent is a real, compounding cost and latency line item, not a free upgrade. Project 2 in this program's real-projects slot (Multi-Agent Research & Report Assistant) is where this module gets applied end to end.