
Doing iconic shit DevRel at @TypesafeAI Queer as fuck
Joined June 2025
- Tweets255
- Following59
- Followers1.6K
- Likes271
I'll be doing a bigger writeup of how I've implemented this, don't worry! But a super short version is the system judges after each word, and Jev has to approve twice in a row because of this case:
"Turn on my dresser lamp..." Jev gets ready
"Turn on my dresser lamp to..." Jev says wait hold up
"Turn on my dresser lamp to 50%..." Jev gets ready
"Turn on my dresser lamp to 50% and..." Jev says EXECUTE
and then the dresser lamp changes while I am finishing saying the other things I want to happen. It's particularly magical to have devices change state one by one as I name them.
As someone who built her first voice-to-control-lights system TWENTY THREE YEARS AGO with a duct-taped microphone on the ceiling that communicated to X-10 lightswitches, Jev allowing me to do something I've never seen before in voice interfaces is such a dream moment.
And on the topic of custom agent harnesses, this writeup of how @sanity_io are using Jev to route to purpose-specific, narrowly-scoped agents is excellent!
Give it a read!
It has been interesting to see all the applications of @typesafeai's Jev at @sanity_io.
Like our new internal agent harness which allows anyone in the company to quickly spin up a bespoke agent. Jev is used for the router and has worked pretty well so far!
We (@rostimelk, @even, and I) wrote up an in-depth post about how it's built so you can do it too.
sanity.io/engineering/how-we…
My favorite Jev use in this system is a "precog engine", where Jev judges the STT as it streams in word by word, warms up intent paths that are determined to be likely, and sometimes fully taking action before I've even finished speaking (depending on risk, reversibility, etc)
I’m on the custom personal agent train for sure, but my particular obsession is extremely responsive and expressive voice interactions. Fully custom STT and TTS pipeline
Jev is enabling so much of it, routing intents to super narrow handler paths that respond lightning fast
Have you ever read a skill with your own eyes that you installed from a vendor? Not like, when you want to pick it apart and make your own, but just using an official skill like Cloudflare's or something
14%Yes all the time
41%Like, once or twice
45%lol never
22 votes • Final resultsLast week it was "set minimum package age to 7 days for everything because the supply-chain apocalypse is here"
This week it's "update all your shit to what was released 5 minutes ago because the 0-day apocalypse is here"
I wonder what next week will be
Allie the Icon retweeted
*on a first date in san francisco*
so are you a queue-by-default or a steer-by-default in the agent kinda person?
T-minus some number of hours before I fork it to start adding Jev powered features
I reallllly don’t want to deal with keeping it in sync with upstream tho
Allie the Icon retweeted
Replying to @allietheicon
the past week or two has informed me that we've barely scratched the surface of understanding the patterns that make jev really shine
The most important thing to do when building with Jev is to keep building with Jev.
The more you rewire your brain to think about software in terms of intelligent primitives, the more you find and revisit ideas you never thought was possible before.
One of my favorite little Jev uses is having it intelligently choose substrings.
In my Agent Tower project, the thread titles are often too long to fit and were truncated. But now Jev picks the best part the string to convey the meaning. Works suuuuuper well
I HAVE BEEN WAITING FOR THIS DAY OMG
I'm fascinated by the role system one models can play in enterprise evals and RL environments...but also in gaming.
Check out @typesafeai's Jev playing @factoriogame. 10 red science for $0.11, playing the actual game (not a gym env).
Allie the Icon retweeted
jev new Box primitive is Yolo for everything
your favorite mug, your cat, your specific plant, it can take all that and more as inputs and find it
should have sent a poet
Today, we are launching OpenRouter Armor in beta.
Tool calls allow agents to interact with the world, and regex-based filtering often isn't enough to detect unwanted tool calls.
Armor uses @typesafeai's Jev model to guard against out of policy tool calls before your agent executes them, with better performance than Claude Code's auto classifier at a fraction of the cost.
Allie the Icon retweeted
Happy JevDay!
But before announcements we have to talk philosophy…
The needs of early adopters / tinkerers / dreamers are quite different from what is needed for extremely stable infrastructure.
For the Jev release, we erred on the latter, but this release is for the former - with a gate so people can opt-in to the danger zone!
Allie the Icon retweeted
what's this?
Jev is evolving!
tune in to our town hall with @CompleteSkeptic and @allietheicon at 1:14 PM to learn more











