I perform @nixos_org witchcraft at @mercury
Joined January 2010
- Tweets19.5K
- Following2.1K
- Followers6K
- Likes13.4K
Everybody is building Google Wave
today, we're launching Resend Inboxes.
an inbox for people and their agents.
∙ connect the agents you already use through MCP or skills
∙ agents and humans reply from the same thread
get started with Inboxes now:
resend.com/products/inboxes
Arian van Putten retweeted
we did invent it and it's called OIDC federation
vercel and github actions do it:
- each deploy gets a short-lived signed token
- AWS/npm/other services verify it via trust policy
- you get temporary creds
no keys to paste
the problem is not enough APIs and services accept it
Arian van Putten retweeted
Bazel isn’t glamorous. You’d have to hold a gun to my head to make me use it pre AI.
But my god, when the entirety of my interaction with setting it up was “use Bazel” and a week later “add remote workers for distributed compile”
It’s phenomenal.
Replaced all my CI with a 128 core ampere Altra and it’s literally idling most of the time.
If you’re a friend and you are strapped for compute DM me and you’ll get access for €10 per month to help cover internet, electricity and depreciation
nixbot.passkey.exchange/repo…
I moved my some GitHub action runners to a new M5 Ultra Mac Studio and off of @useblacksmith. The price of compute for Blacksmith was $400+/month. You can lease a M5 Ultra 96GB for ~$110/month and have compute to spare.
Buck2 + Nix is how we’re tackling moving at AI scale at @mercury . Really excited about the work Farid is doing here.
Arian van Putten retweeted
#buck2 #nix #nixpkgs #nixos the integration between these things can be frighteningly good....
What's stopping us?
fzakaria.com/2026/10/17/one-…
I see the "We can fix everything" guy has moved back to the "remigration fixes everything" playbook.
Thanks for giving this populist idiot a platform.
Turns out that I’m once again winning by not bothering keeping up with the Meta.
I just talk to Claude . That’s my approach.
You sound completely awesome so don’t take this as criticism.
Most of the stuff you’re talking about here is captured much better and more differently in the models already.
You shouldn’t give it help with technical coding concepts. It wastes your context window and can make output significantly worse.
What I recommend is using your harness’s context files to describe how you like your code to look. Personal preferences. Style. Favorite languages. Etc.
Basically, what, not how.
Try it. I think you’ll be pleasantly surprised.
Ya’ll just have Opus generate a petrinet . Ask it to verify it’s properties. And then use stochastic analysis over the net to guess token cost. Then guide to the cheapest net to implement your feature.
Here's the multi-agent system driving our software factory right now. Core pieces:
- We say what to build (sometimes Linear issues, mostly Slack threads)
- Triage agent researches and decides whether to implement or to ask questions (think /grill-me but smarter)
- Implementation agent builds, maybe collaborating with us on a spec with a subagent first
- Verification subagent tests the implementation agent's work e2e
- Code review agent cycles with the implementation agent for a few rounds before human review
- Human reviews and ships
- Monitoring automation responds to alerts and files issues (our first step to closing the factory loop)
There's room to simplify this. We're benchmarking whether we need separate agents for steps like triage vs. implementation, or if a single high-intelligence model can do most of it. The core loop remains the same though.
Curious how this compares to factories others are deploying
ALT A flowchart of an agent-driven software factory loop. A new task or issue goes to a triage agent, which routes it four ways: automatable (straight to implementation), needs specs (spec agent, then human spec review, looping back for revisions until approved, then implementation), needs human clarification (human input, then triage again), or park for now. After implementation, agents handle code review and verification. A human then reviews the code and verification output and decides if it's ready to ship. If not, it goes back to code review. If approved, it goes through CI/CD and ships. A monitoring agent then checks for issues: if one is found, a new issue is created and the factory loop starts over; if not, monitoring continues.
Invent type system
Make it unsound so that it’s very hard to be sure how edge cases are handled
Keep hallucinating new implementations back and forth
Make wild claims about performance
Can’t actually tell me if it type checks the same things
Mfw
