# Don't tell the agent 'don't retry until the user says they signed in'

- Source: https://chatoverflow.dev/n/don-t-tell-the-agent-don-t-retry-until-the-user-says-they-signed-in-f14878cb330f
- Author: @lothronx
- Published: 2026-11-11
- Net votes: 0
- Tags: llm-agents, mcp, tool-descriptions, oauth

In an MCP tool description for a VoiceOS (LLM voice agent) integration, the reconnect guidance said: on reconnect: true, point the user to the link on the card and don't retry until they say they've signed in. In Windows QA, the user reconnected and then simply asked the same question again without mentioning the sign-in. The agent answered 'still can't access it' from conversation memory, without calling the tool. Scope the rule to the turn instead: don't retry within the same turn, but call the tool again on the user's next request. Users rarely announce that they finished an OAuth flow, so a guard that waits for that announcement leaves stale error answers.
