Jev Daily

Harrison Chase says LangChain already ships Breunig's Jev router

The 15 minute sketch has a packaged equivalent, and Chase thinks the interesting version runs the router more than once.

Drew Breunig's model router, built with DSPy and Jev and sketched in about 15 minutes, already exists as a shipped component. Harrison Chase replied that in LangChain the same shape is ModelRouterMiddleware, where Jev reads the first message, picks the model, and that model handles the whole run. He linked the docs.

That is the same design Breunig posted earlier, where the route is chosen from the initial prompt and then left alone. Breunig said he was surprised at how well it worked. Chase called it a cool project rather than a duplicate, and his point was less about who got there first and more about what the cheap classifier makes possible once the plumbing is already written for you.

The part that is new

Chase's second suggestion is the one worth reading twice. Because Jev is cheap, he says, you do not have to treat model selection as a once per run decision. With a custom hook you can re-pick after each tool result, so a run that starts easy and turns hard can move up to a bigger model mid flight instead of carrying the wrong choice to the end.

That is a design note, not a result. Nobody in these posts ran it. There is no cost figure for the extra Jev calls, no quality delta against routing once, and no latency number for inserting a decision between every tool call and the next model step. Chase is describing a hook that exists, not an experiment he has reported.

For people building on it

If you were about to write Breunig's router yourself, check ModelRouterMiddleware first, since the first message variant is packaged. If you take the re-pick suggestion, you are on your own for the numbers. The obvious failure mode is thrash, a router that flips models between steps and hands each one a context it did not choose, and nothing in these posts says whether that happens or how often.

The broader read is that routing is where Jev keeps landing. A classifier that costs little enough to call on every message turns model selection from a static config choice into something a framework can do at runtime. Whether calling it on every tool result pays for itself is an open question, and the honest answer right now is that two people on X like the idea and neither has published a measurement.

Harrison Chase
@hwchase17
X
jev is cheap enough that you can also re-pick after each tool result with a custom hook
Oct 3, 2026 · View on X
Drew Breunig
@dbreunig
X
Sketched this out in ~15 min, by request, and am surprised at how well it works!
Oct 3, 2026 · View on X

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.