ATLAS Exposes the Search Weakness of AI Agents

🔥 What happened Exa launched ATLAS, a new benchmark of 547 real-world research tasks grading agentic web search. No agent costing under $1 per task clears a row F1 of 0.5. 💡 Why it matters Even the priciest search agents miss roughly a third of golden results. Swapping search backends moves scores by 16 points with the model harness held fixed — the search layer, not the model, is now the bottleneck. ⚡ Our take If you're still tuning models in 2026 while ignoring your search stack, you're optimizing the wrong half of the system. Exa conveniently placed itself on the Pareto frontier with its own benchmark — elegant marketing, but rivals now have to disprove it.