Metaview shipped Jev into every agent on its platform
Shahriar Tajbakhsh reports minutes to seconds on candidate search, with no baseline numbers posted.
Shahriar Tajbakhsh said Metaview shipped Jev into every agent on the platform over the weekend. His headline result is that candidate searches in the sourcing product that used to take minutes now take seconds, at what he describes as the same accuracy, roughly 10x faster, and meaningfully cheaper per search.
candidate searches in our sourcing product that took minutes now take seconds
That is the whole of the measurement. There is no baseline latency in seconds, no accuracy figure on either side, no cost per search before or after, and no description of the evaluation set the accuracy claim is based on. It is a team reporting on its own production workload after a weekend of work, which is worth knowing and is not a benchmark.
The part Tajbakhsh cares about
He says the speedup is not the interesting part. The change he describes is structural. Metaview broke each agent into what he calls its smallest semantic units, queried each one separately, set its own thresholds on the results, and now fixes bugs by adding a question rather than by rewriting a system prompt. He calls that a different way to build and says it is now the company's baseline.
That is the pattern worth taking away for anyone deciding how much of a product to hang on Jev. The claim is not that a model got smarter, it is that a class of decisions got cheap enough to make many small ones instead of one large one, with thresholds the team controls rather than behaviour they negotiate with a prompt. Whether that holds up outside recruiting search is untested here.
Tajbakhsh credited the TypeSafe team and closed with a line about optimizing for intelligence per dollar having unintuitive consequences, tagging Diogo Almeida, who replied that the industry is going full Jevons faster than anyone ever thought.
the industry is going full jevons faster than anyone ever thought
For people building on it
Treat the 10x as Metaview's number, not a measured one. The direction matches other reports of Jev cutting latency inside an existing agent loop, including Kyle Jeong's Stagehand Act integration, but two separately posted results from two teams on two different workloads are not a trend and are not comparable to each other. If you are planning the same refactor, the thing to copy is the decomposition into small queries with explicit thresholds, and the thing to measure yourself is whether accuracy really holds when you split one prompt into many questions. Metaview says it did on their sourcing searches. Nobody outside Metaview can check that yet.