Sanity uses Jev as the router in its internal agent harness

Sanity has an internal agent harness that lets anyone at the company spin up a bespoke agent, and Jev is the piece doing the routing. knut, who posted the news, says it has worked pretty well so far, and that he wrote up how the thing is built together with @rostimelk and @even so other teams can copy the structure.

knut

@kmelve

X

Jev is used for the router and has worked pretty well so far!

That is the whole of what has been shared. No accuracy number, no latency figure, no cost per run, no count of how many agents people inside the company have actually spun up. The claim is an impression from the team that built it, not a measurement, and the post is explicit that the point of the write-up is reproducibility rather than a result.

What the router job actually is

Routing in a harness like this means deciding which agent, tool or path a given request goes to before the expensive model does any work. It is a classification problem with a small output space, which is why it keeps showing up as the slot people fill with a cheap, fast model rather than a frontier one. A router that quietly picks wrong does not throw an error, it sends the request somewhere plausible and the cost shows up downstream as a bad answer, so a team's sense that it is working fine is weaker evidence here than it would be for a generation task.

For people building on it

Treat the Sanity post as a shape to borrow, not as evidence that Jev routes well. If you lift the pattern, the number you want and do not get from this announcement is agreement between the router's choice and the choice a human would have made on your own traffic, measured on a sample you label yourself. Internal harnesses are also a forgiving environment. Everyone using it works at the same company, knows what the agents are for, and can shrug off a misroute by trying again, which is not true of the same router sitting in front of customer traffic.