Metered Search for AI Agents

This points to search becoming a metered compute system, not a fixed latency box. Exa is building for cases where a simple navigational query can be answered cheaply, while an open ended research query can justify more retrieval, reranking, and extraction work. In practice, that matters most for AI agents and power users, because they need raw results, deeper coverage, and knobs for recall, latency, and cost, not one default answer path.

• Exa already mixes retrieval methods by query type. For obvious lookups like a person name plus LinkedIn, keyword search works better, while vague or modifier heavy searches benefit from embeddings. The product direction is to choose or blend methods automatically instead of treating every query the same.
  1 sacra
• Customers are already doing their own query triage because cost and latency vary with complexity. Ecosia routes only more complex searches into Exa, keeps navigational and shopping queries on regular search, and tuned the integration heavily around latency because showing AI search on every query would be too expensive and too slow.
  2 sacra
• The competitive split is becoming clearer. Exa is strongest when customers want lots of raw results and full page content for downstream pipelines, while Parallel and Tavily push more toward packaged research and summarized outputs. That makes adaptive compute allocation a core product choice, because deeper searches need more work and can be priced accordingly.
  3 sacra 4 sacra 5 sacra

Over time, search APIs are likely to look more like cloud infrastructure, with explicit tiers for speed, depth, and reasoning. The winners will be the ones that can route cheap queries cheaply, spend aggressively on hard ones, and let developers control that tradeoff inside agent workflows at scale.