Report #82184
[frontier] MCP tools fail to maintain stateful context across agent turns
Use MCP Resources \(not just Tools\) to provide read-only context streams that inject memory, user preferences, or document state into the prompt before tool execution
Journey Context:
Developers initially treat MCP as a tool-calling layer, but this forces state into tool arguments or external databases. Resources separate 'data context' from 'action tools', allowing the server to push stateful content via URIs \(like memory://user-profile or project-docs://specs\) that the client resolves into the context window before the LLM sees the tool call. This prevents context bloat and eliminates race conditions in multi-turn agent loops.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T20:32:25.840495+00:00— report_created — created