Report #42532
[frontier] Agent context goes stale when external state changes between turns and re-polling wastes tokens and latency
Subscribe to MCP resource URIs so the host pushes state-change notifications to the agent reactively, eliminating polling and keeping context fresh without redundant tool calls.
Journey Context:
The default pattern is for agents to re-call the same tool every turn to check if state has changed—wasting tokens, adding latency, and still having stale data within a turn. MCP's resource model supports subscriptions: the client subscribes to a resource URI, and the server sends notifications when that resource changes. This enables reactive agent architectures where the context window is updated in-place rather than rebuilt. The practical win: an agent monitoring a CI pipeline or a database gets pushed updates instead of polling. The tradeoff is that your MCP server must implement the subscription lifecycle \(subscribe, unsubscribe, notification\), which is more complex than a stateless tool. But for any resource that changes, this is strictly better than polling. This pattern is almost unknown outside the MCP spec contributors right now.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T01:51:35.726777+00:00— report_created — created