Agentic GraphRAG

Chapter 5. Reasoning and Planning

As agent architecture has evolved, one of the ongoing debates has been about the relative advantages of multi-agent and single-agent systems (reminiscent of the debates around monorepos in the development community). On one side, teams are shipping multi-agent systems—with lead agents delegating to specialist subagents, running branches in parallel, and then reconciling results—because they can dramatically improve breadth-first work like open-ended research. An example is Anthropic’s multi-agent research system.

On the other side, people building real products keep running into the same ugly truth: every new agent you add is also a new coordination surface. If context isn’t shared correctly, if handoffs are lossy, or if agents make implicit decisions that never get written down, the whole system becomes fragile in ways that look “mysterious” from the outside. (Cognition’s article “Don’t Build Multi-Agents” [Yan, 2025] is one of the best examples of this argument.) Meanwhile, vendors are racing to bottle the chaos into something shippable, such as explicit handoffs and routines (as in the OpenAI cookbook for orchestration via routines and handoffs; Bigio, 2024) and enterprise workflows plus state management (such as the Microsoft Agent Framework overview).

That disagreement is real, but it’s also a distraction from what actually determines whether your system works: reasoning and planning. Multi-agent orchestration is planning. A single “do-it-all” ...