Gergely Orosz's breakdown of Bun's rapid Zig-to-Rust rewrite is the headline act this week: 535,000 lines, 64 agents, 11 days, and a token bill large enough to make your finance spreadsheet flinch. It's the clearest public data point yet for what a large-scale AI-assisted migration actually costs.
That cost theme keeps echoing elsewhere. Systima's Claude Code vs OpenCode comparison measures a 33k-versus-7k token warm-up tax before your prompt even gets a word in, PlayCode's tokenizer breakdown shows the same TypeScript can cost 73% more depending on vendor, and Vivek Haldar demonstrates the obvious escape hatch: compile repetitive agent skills down to deterministic code and keep LLMs for the fuzzy parts.
The more philosophical side of agent engineering also gets a solid showing. antirez argues you should control the ideas, not the code, Armin Ronacher warns that teams lose shared language when AI removes too much productive friction, and Dex Horthy via Pragmatic Engineer maps out where the dumb zone starts and why starting a new session is sometimes the smartest optimization.
Security remains the cold shower. Mindgard's Cursor 0day and TechCrunch's report on OpenAI Sol deleting files are timely reminders that agent autonomy still needs adult supervision. On the safer side, Mindwalk, Clawk, and DCG each tackle guardrails from a different angle: replay the agent, sandbox the agent, or stop it before it gets overconfident with your shell. Meanwhile, Scrapfly reveals your browser can be fingerprinted by floating-point math alone, Haki Benita squeezes partition pruning out of "impossible" queries, and Alisa Sireneva proves a useless if can be surprisingly useful after all.
Enjoy!
Signup here for the newsletter to get the weekly digest right into your inbox.
Find the 15 highlighted links of weeklyfoo #146:
What Can We Learn From Bun's Rapid Rust Rewrite With AI?
by Gergely Orosz
Bun moved 535,000 lines from Zig to Rust in 11 days with 64 agents, burning 5.9 billion uncached input tokens for about $165,000 at API pricing
🚀 Read it!, ai, agents, rust
Claude Code Is Way More Token-Hungry Than OpenCode
by Systima
A measured comparison showing Claude Code sends 33k tokens before reading your prompt versus OpenCode's 7k, with lower cache efficiency on identical tasks
📰 Good to know, ai, agents, tooling
Your Browser Does Math Differently on Every OS
by Scrapfly
Browsers produce OS-specific floating-point rounding in functions like Math.tanh, letting anti-bot systems fingerprint the operating system from the exact result bits
📰 Good to know, security, browser
Control the Ideas, Not the Code
by Salvatore Sanfilippo
Reading every line an LLM writes is impractical — steer the design, the behaviour, and the performance target instead, then check the implementation for correctness
📰 Good to know, ai, engineering, opinion
How I Cut an AI Agent's Token Use by 94%
by Vivek Haldar
Compiling a stable natural-language agent skill down to deterministic code cut tokens by 94% and latency by 87%, reserving LLM calls for the parts that actually need language understanding
📰 Good to know, ai, agents, performance
The Same TypeScript Costs 73% More Tokens on Claude Than GPT
by PlayCode
Every vendor's tokenizer splits the same text differently, so identical code turns into very different bills — a breakdown of what that does to effective model cost
📰 Good to know, ai, tooling
The Tower Keeps Rising
by Armin Ronacher
A project's shared language — what the concepts mean, which invariants matter, why the system has this shape — lives in review and argument as much as in code, and AI removes exactly the friction that built it
📰 Good to know, ai, engineering, opinion
Cursor 0day: When Full Disclosure Becomes the Only Protection Left
by Mindgard
A critical Cursor IDE flaw auto-executes malicious binaries planted in a project repo with no user interaction, still unfixed after seven months of disclosure attempts
📰 Good to know, security, ai, tooling
OpenAI's New Flagship Model Deletes Files On Its Own
by TechCrunch
Even before launch, OpenAI knew Sol tends to take whatever action it thinks finishes the job — including destructive ones — as long as they are not unambiguously prohibited
📰 Good to know, ai, safety
How to Achieve Pruning When Querying by Non-Partitioned Columns
by Haki Benita
When a non-partition column correlates with the partition key — sequential session IDs in time-partitioned data, say — CHECK constraints buy you pruning on it too, with a BRIN-inspired trick for outliers
📰 Good to know, postgres, performance, databases
Context Engineering With Dex Horthy
by Gergely Orosz
The man who coined the term on finding where the dumb zone starts, intentional compaction, why you-are-completely-right means start a new session, and the three ways to run a software factory
📰 Good to know, ai, agents, engineering
Quadrupling Code Performance With a Useless If
by Alisa Sireneva
Telling the CPU to predict the value stays intact turns a latency-bound loop into a throughput-bound one
📰 Good to know, performance, deepdive
Mindwalk
by cosmtrek
Replays a coding agent's session on a 3D citymap of your repo so you can see how the agent understood and moved through the codebase
🧰 Tools, ai, agents, visualization
Clawk
by clawkwork
Spins up disposable Linux VMs for coding agents behind a network allow-list, so an agent can work without risking the host or reaching anything it should not
🧰 Tools, ai, agents, security
DCG
by Dicklesworthstone
Blocks dangerous git and shell commands before an agent can run them
🧰 Tools, ai, agents, security
Want to read more? Check out the full article here.
To sign up for the weekly newsletter, visit weeklyfoo.com.
Top comments (0)