Agents require completely different search inputs, outputs, and latencies “The problem’s inputs are different, outputs are different, and constraints are different. Imagine someone running an agent built with a Luna model and someone running an agent built with a Fable model. They are very different models. How you want... to optimize signal-to-noise and tokens for each of them is so different in terms of what you do in the web search stack.” Parag Agrawal Love to hear your thoughts Sarah Sachs Simon Last Lance Martin Charles Packershow more

The useful unit is the task, not the search result. For an insurance claim, 20 relevant pages can still be worse than one policy clause with its version and source. I'd benchmark retrieval on whether the agent makes the right decision, not just how fast it returns tokens.

for small saas it means your docs are the new landing page as an agent rarely sees the hero

so the search bar needs a passport for every agent

so tune the stack per model. sounds expensive

👍 the larger the model generally the better it will be at handling messy context windows, allowing you to be lazier with the harness (and lazier w/ how you do retrieval)

agentic search is part of the model stack, not a generic infrastructure layer. the winning systems will optimise retrieval according to the agent's behaviour, not just relevance.

that changes what “better” even means. a quick agent that returns a plausible answer is great until the job is actually research and the missing context is the whole cost.

the ai tagging is smart engagement bait btw

Agent search will fragment by job: research wants provenance and breadth; commerce wants structured inventory and price; coding wants low latency and exact context. A single search API is unlikely to serve every agent well.

does the search api need to know which model is calling it, or can the agent set its own token/noise budget? swapping models mid-task seems like the awkward case