The standout this week arrives courtesy of mathematics: Kevin Buzzard's post documents AI tools now writing Lean proofs for theorems they discovered themselves — the formalization feedback loop has inverted, and human mathematicians are running to keep up. Not a hypothetical; it's happening now.
Agent economics dominate the rest of the issue. Cursor's deep dive into planner-plus-worker swarms shows they beat a single frontier model on speed, quality, and cost. Stencil's prewalk pattern gives the approach a name and a workflow: have the frontier model explore and dump its understanding to a todo list, then hand execution to a cheaper model. Adam Jacob goes a step further and shows how structuring agent tasks into deterministic code cut token usage 8x — real before-and-after numbers, not vibes. Netflix spent the same energy differently: benchmarking vLLM and Triton to build their own serving stack rather than paying hosted API bills. And Google is going further still: fusing Gemini's architecture directly into silicon, so the hardware runs one model by design.
One finding deserves its own paragraph: Greptile's study shows Claude Code and Codex each catch more bugs in the other model's output than in their own. Cross-model routing for code review is no longer just a fun idea — it's the obvious call.
Security delivered a string of memorable entries. Elastic's report on the Contagious Interview campaign hides malware inside SVG images inside a fake coding-interview challenge — targeted, specific, and realistic for any JavaScript developer. Socket.dev catalogued 53 slopsquatting targets across five popular LLMs: the fake npm package names models keep hallucinating, mapped to how fast attackers are registering them. And Matteo Collina lays out why you cannot harden Node core against prototype pollution — the fix belongs at the boundary, not the runtime.
Rounding out the issue: Fetch gets typed error codes thanks to a TC39 proposal that would end the era of every network failure collapsing into a bare TypeError. Arshad Yaseen shows data-oriented design producing a JavaScript parser faster than Oxc, SWC, and Babel, with the full engineering story. Ben Dicken's animated explainer on routing queries across a petabyte of Postgres presented as a single database. Sunil Pai frames a UX problem nobody has solved yet: what does it look like when an agent is editing the same document you are, simultaneously? And Oxc type-aware linting lands stable — unhandled promises, unsafe assignments, all in Rust.
On the tools front: LoopGain cuts agent loops the moment iterations stop improving (someone had to build this), Claude Thermos keeps the prompt cache warm through long multi-subagent sessions so you don't get re-billed for re-encoding, A11Y.md hands AI agents enforceable WCAG-aligned accessibility rules as a context file, and Shadscan runs nearly 60 deterministic accessibility and composition checks on your shadcn components.
Enjoy!
Signup here for the newsletter to get the weekly digest right into your inbox.
Find the 19 highlighted links of weeklyfoo #147:
In-House LLM Serving at Netflix
by Netflix
How Netflix built an in-house LLM serving platform on vLLM and Triton instead of hosted APIs
📰 Good to know, ai,infra
Fetch Needs Error Codes
by James Snell
Fetch collapses every network failure into a bare TypeError, and a TC39 proposal would thread typed error codes through
📰 Good to know, javascript,web
No, We Can't Harden Node.js Against Prototype Pollution
by Matteo Collina
Matteo Collina argues you cannot harden Node core against prototype pollution and the fix belongs at the boundary
📰 Good to know, security,nodejs
A Practical Guide to Reducing Token Spend
by Adam Jacob
A swamp workflow structures agent tasks into deterministic code and cut token usage 8x
📰 Good to know, ai,agents
Agent Swarms and the New Model Economics
by Cursor
Planner agents decompose work and cheaper workers execute it, beating a single frontier model on speed, quality and cost
📰 Good to know, ai,agents
You Only Need the Frontier Model for One Single Edit
by Stencil
Have the frontier model explore and dump its understanding to a todo list before handing execution to a cheaper model
📰 Good to know, ai,agents
Google Is Building a Chip With Gemini Baked Into the Silicon
by The Next Web
Fusing one model's architecture into the circuitry instead of running any model on general hardware
📰 Good to know, ai,hardware
Human Mathematicians Are Being Outcounterexampled
by Kevin Buzzard
AI tools now solve theorems and write Lean proofs to prove their own results, speeding up formalization research
📰 Good to know, ai,math
One Document, Two Hands
by Sunil Pai
Designing for the case where an agent is a guest in your app, editing the same document you are
📰 Good to know, ai,agents
Engineering High-Performance Parsers with Data-Oriented Design
by Arshad Yaseen
How data-oriented design built Yuku, a JavaScript parser faster than Oxc, SWC or Babel
📰 Good to know, javascript,performance
How a Fake Interview's Coding Challenge Steals Credentials
by Elastic
A North Korean campaign hides malware in SVG images inside a fake job interview JavaScript coding challenge
📰 Good to know, security
Models Are Worse at Reviewing Their Own Code
by Greptile
Claude Code and Codex each catch more bugs in the other model's code than their own, so route reviews across models
📰 Good to know, ai,agents
Making 768 Servers Look Like 1
by Ben Dicken
An animated explainer of how a petabyte of Postgres across 256 shards presents as one database
📰 Good to know, database
Oxc Type-Aware Linting Is Stable
by Oxc
Rules relying on the TypeScript type system, like unhandled promises and unsafe assignments, now land in the Rust linter
📰 Good to know, javascript,tools
53 Slopsquatting Targets Across 5 Popular LLMs
by Sarah Gooding
A study of the fake npm package names LLMs keep inventing, and how attackers register them
📰 Good to know, security,ai
LoopGain
by LoopGain
Open-source tool that stops AI agent loops once iterations no longer improve
🧰 Tools, ai,agents,tools
A11Y.md
by Felipe Arriaga Carrico
Context system giving AI agents enforceable accessibility rules aligned with WCAG
🧰 Tools, ai,tools
Claude Thermos
by Ivan Zeigerman
Keeps the prompt cache warm during long agent sessions so waiting on subagents doesn't re-bill re-encoding
🧰 Tools, ai,agents,tools
Shadscan
by OrcDev
Static analysis for shadcn components with almost 60 deterministic accessibility, state and composition checks
🧰 Tools, react,tools
Want to read more? Check out the full article here.
To sign up for the weekly newsletter, visit weeklyfoo.com.
Top comments (0)