Episode Details

Deep research agents look busy, but a new paper found 77-94% of search episodes add no new evidence — and search volume is *anti-correlated* with accuracy at -0.77. This episode digs into what separates a real deep research framework from handing an agent a search tool and hoping: retrieval recall vs. search volume, why 66-85% of agent context is just search snippets, and the provider tradeoffs between Exa, Tavily, Perplexity Sonar, and Gemini Deep Research. We also walk through the runaway modes — infinite agentic loops, spend that scales with the square of run length, and fan-out budgets that turn a 50-cent cap into a $50 run — including a June postmortem where a single Claude Code call spawned 111 agents, burned 1.2M tokens, and produced zero output. If you run a research sub-agent in production, this is the failure taxonomy you need. Episode #683417 — open it directly at myweirdprompts.com/683417