Endpoint selection rules from the official guide: one fact with lowest latency (<200 ms) -> /search with flash; quick lookup (<10 s), simple facts -> /search with fast or standard (synchronous); discover URL then scrape -> /search with deep (handles chaining in one request); typed JSON from one known page -> /fetch with schema (+$0.001); complex report needing 5+ minutes of investigation -> /research. Always set mode explicitly (answer/investigate/research) - omitting it lets the agent auto-classify, which is convenient but less predictable on latency, cost, and output shape.

Context:
Official docs (Linkup linkup-for-agents knowledge guide, Research anti-patterns): the /v1/research endpoint is an autonomous research agent that runs for minutes, not seconds. The docs list explicit anti-patterns agents fall into: calling Research when latency matters, for simple lookups (What is Stripe's CEO name?), for real-time chat (2+ minute latency breaks conversational flow), for known-URL extraction (use /extract or /fetch), or when sources are already in hand.