From when I first started using @nextjs I always found @vercel to be absolute best in class for dx
Building this slackbot was incredibly simple with @aisdk and @chatsdk, I didn’t read the slack docs once
Always bet on TypeScript and always bet on @vercel
At the end of @usekernel ‘s monthly retro, I always end the meeting with, “Let’s have a great month.”
It comes from the restaurant world, wherein the chef de cuisine starts every night with a rundown of who’s coming, what’s cooking—ending with, “Let’s have a great service.”
The best teams operate like the best kitchens: high energy, team camaraderie, in service of the customer. I live for this feeling.
Conductor has the distinct edge of being an IDE born for coding agents. An ADE if you will.
Agents will take remote dev mainstream. At most large and sensitive organizations, “local” isn’t a thing. But this future wasn’t evenly distributed.
Excited to power their Sandbox!
Conductor's parallel coding agents were local-only, but now they can run remotely on Vercel.
"Our customers can't tell the difference because Vercel's Sandboxes are so fast." https://t.co/ItuYl5F331
we shipped an "mcp app" so tool calls render with proper uis. pretty neat!
most importantly, this means diffs render properly (not just as an unreadable json blob)
also has (limited) html previews, logs, and more!
there's a ton of demand right now for excellent infra talent, definitely outpacing supply
im hiring engineers who have tackled megascale, multicloud product infra to work on the Claude Platform. help me find them 🙂
Every VC had a reason Vercel shouldn't exist.
Guillermo Rauch heard a different reason from every investor he spoke to when building Vercel, and he kept going anyway.
He joined us live at Mercury HQ in SF for Founders in Arms, and this was one of the sharpest moments in the conversation: the idea that consensus is a signal to walk away, not lean in.
Full conversation with Guillermo Rauch (CEO, Vercel) on the latest Founders in Arms episode. Link in bio.
I made a fax machine for your agents!
~~ Introducing agentfax ~~
Why?
This morning, I had to fax the IRS. Claude Code got me 90% of the way there – even filling out the PDF! – but it couldn't actually send the fax
So I had Claude Code find a fax API, wrap it in a @tempo proxy, and deployed on Val Town. Took me two hours end-to-end (while working on other stuff)
I wonder if anyone will use it! Does anyone fax anymore?
Why don't LLMs run their code more often?
One of the first things I learned as an engineer was to run my code as early and often as possible:
Step 1. Get *something* running.
Step 2. Get it deployed.
Step 3. Make a change. Confirm it runs. Repeat.
Step 4. Send a PR / Deploy.
At every point, your code runs or you get fast feedback on what broke it.
LLMs – even in premier harnesses – often do the novice thing of writing thousands of lines of code before running & deploying. Sometimes they'll even confidently tell you they're done without ever testing any running code!
Why? LLMs are writing professional code, but iterating like a freshman in CS 101.
My guess is that they avoid running code, because it's so expensive relative to just outputting more code.
In order to run code, an LLM has to do another turn through its tool loop. I think this is why agents tend to write a lot, then verify once at the end – or never.
But for us humans, we've taken pains to optimize the edit + run flow into a single command, like cmd+enter, or just hitting save and hot reloading. This way, verification is free, instant, and woven into the loop. Why shouldn't it be the same for LLMs – a single tool call that makes the edit *and* runs it?
I was surprised I couldn't find more people talking about this pattern. Aider, Cursor, Cline, Replit, Copilot all have "auto run after edit," but at the harness level: the LLM still issues a separate tool call and still pays the round trip, which is expensive and slow.
Today, we shipped a small tweak to Val Town's MCP which should fix this. Now our edit file tools take an optional `run` object alongside any edit, so the LLM can test the change immediately after it's applied:
{
"tool": "replace_in_file",
"input": {
"val": "stevekrouse/my-api",
"path": "main.ts",
"old_str": "return 'hello'",
"new_str": "return 'world'",
"run": {
"kind": "fetch_val_endpoint",
"method": "GET" ,
"path": "main.ts"
}
}
}
I think this works for Val Town particularly well, because our deploys are 100ms, but I'm still confused why I haven't seen more of this pattern elsewhere.
If you've seen prior art on fused edit+run tools, please send!
The main alternative I considered is Cloudflare's code mode, which would be more composable and elegant. Instead of manually combining two tools like this, code mode lets LLMs compose tools itself in TypeScript. However, I decided to go ahead with this little experiment despite the "smell" because I think edit+run is a fundamental enough pattern to enable fast feedback loops that it was worth it