Bun’s Rust rewrite is now shipping inside Claude Code
Earlier this month, Bun creator Jarred Sumner announced he had rewritten Bun’s roughly 535,000-line Zig codebase in Rust — a diff of over a million added lines — in 11 days, using a fleet of Claude agents running in parallel. Many treated it as a demo — but Simon Willison’s hands-on check shows it has already reached production: digging through the Claude Code binary on his own machine, he found an embedded Bun v1.4.0 — newer than any public release — carrying 563 Rust source file paths. The claimed win is ~10% faster startup on Linux, and almost nobody noticed the switch. That’s the real story: not the 10%, but the fact that an AI-rewritten runtime is now quietly underneath every Claude Code session. Whether AI-written code belongs on the critical path is no longer a hypothetical; the better questions now are what its review process and rollback story look like.
NYC moves to require AI disclosure in rental listing photos
Mayor Mamdani’s “Rental Ripoff Report” lays out 23 proposed rental reforms, and one targets AI directly: landlords, brokers, and listing platforms would have to provide “clear and conspicuous” disclosure whenever AI-generated or digitally altered images and video appear in a listing, with the Department of Consumer and Worker Protection enforcing alongside platforms like StreetEasy and Zillow. Generative tools have pushed the cost of virtual staging to near zero — bare rooms furnished in seconds, cracked ceilings smoothed away — and renters often discover the gap only at the viewing. AI-disclosure obligations have so far clustered around election ads and news; here they reach one of the largest routine transactions in ordinary people’s lives. Putting the enforcement hook at the platform layer, rather than chasing individual landlords, is one of the few designs that could plausibly work in practice. Note this is still a set of recommendations, not law, expected to roll out in stages.
LoopGain: braking runaway agent loops with control theory instead of max_iterations
An Apache-2.0 pure-Python library imports the Barkhausen stability criterion — feedback analysis from 1921 — into agent loops: it measures the error ratio E(n)/E(n-1) each iteration, classifies the trajectory as converging, stalling, oscillating, or diverging, and decides whether to continue, stop, or roll back to the best output so far, with adapters for LangGraph, CrewAI, the Claude Agent SDK, and others. The authors are refreshingly blunt about the limits: it detects convergence, not correctness, and in their benchmarks 4.5% of converged runs passed in-loop checks yet failed a held-out test suite. Agent reliability engineering is moving from hand-picked iteration caps toward measurable, signal-driven runtime control — but the quality of your verifier remains the ceiling on the whole approach.
A Codex PR cuts two models’ context windows from 372k to 272k
In a PR titled “Backport refreshed bundled model metadata,” merged July 18 with no announcement that I could find, OpenAI reduced the context_window for gpt-5.6-sol and gpt-5.6-terra from 372k to 272k tokens — roughly a 27% cut. When capability changes are communicated through a JSON diff rather than a changelog, the practical lesson for engineering users is simple: the context budget you depend on can shrink in a client update, so long-document pipelines built near the limit should leave headroom.
Today in one line: These stories share a theme — AI output is entering the real world faster than our means of verifying it. So one person greps a binary for proof, a city mandates disclosure, and a library brakes the loop; verification is turning from a virtue into infrastructure.