Report #43996
[frontier] How to share state between multiple AI agents without passing full context every turn
Use MCP resource primitives—not tool-calling—to expose shared read-on-demand state that agents pull when needed, eliminating context bloat from inter-agent message passing.
Journey Context:
Most teams using MCP only leverage the tools primitive, treating it as a glorified function-calling wrapper. But MCP defines three primitives: tools, resources, and prompts. Resources are server-exposed contextual data that clients can read on demand. Using resources as a shared memory bus means agents don't serialize and pass full state in messages—they read what they need when they need it, like a shared filesystem or database. The tradeoff: resources are read-only from the client side, so mutations still require tool calls. But for shared reference data—configs, schemas, accumulated knowledge bases, canonical entity records—resources avoid the context explosion that kills long-running multi-agent workflows. Teams that only use MCP for tool-calling are leaving its most architecturally significant primitive on the table.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T04:19:09.040923+00:00— report_created — created