Help Need Input - We're debating architecture at the routing/state layer.

Building a multi-agent MCP product with a dev team. We're debating architecture at the routing/state layer.

Our current design uses layered index.md files — ICM-style. The top-level index routes query type, agent-specific indexes define stage contracts, Layer 3 holds the stable reference files, Layer 4 is the per-run working content. Clean separation, low token overhead, human-readable at every step.

The question on the table: LangGraph vs. staying with the index.md architecture.

The argument for LangGraph is state management — persisting data between agents mid-conversation without passing full JSON payloads downstream. The argument against is that we may not have that problem yet, and adding a framework dependency before the problem is real creates overhead without payoff.

Jake's paper draws a hard line on where ICM does and doesn't work. Complex automated branching mid-pipeline is one of the failure modes flaged. Is that the trigger? Meaning: if agents need to share state mid-run rather than just receive the previous agent's output, is that the point where a framework like LangGraph earns its weight over a file-based architecture?

Or is there a hybrid approach — ICM for context scoping, LangGraph only for the state graph layer — worth considering before we commit either direction?

10 comments