A trojan Claude Code extension in Cursor's marketplace, Roblox's 70% AI review acceptance, and Anthropic drops the five-hour cap

Roblox — mines years of human code-review feedback into rules, called exemplars, that steer both agent-written code and AI review comments. Its AI-generated review comments are accepted nearly 70% of the time, against 50–55% for human-written ones, and about 5% of PRs now merge automatically after AI review, safety checks and a risk score. The pattern to copy: engineers approve the rules, not every diff. VP of Engineering Andrew Swerdlow says caching and shared context cut the cost per rule evaluation by 90%. Engineering Enablement by DX

Proton — cancelled company-wide AI 're-interviews' — sessions where staff would have had to justify their AI use, with dismissal implied — a day after 404 Media reported on his aggressive AI push, brushed-aside security concerns and a culture of fear described by employees. The canonical case of an AI mandate done badly; expect your team to raise it. The same measures will move into performance reviews instead, and Yen says the AI push continues. 404 Media

Cisco — gave a 27-year-old engineer, hired through the EzDubs acquisition, unlimited AI tokens to refactor legacy code with a two-person team — a four-week proof first, then a 24-month goal covering 98% of the legacy estate. A senior-leadership bet that AI-native early-career hires can hold work once gated on decades of experience. About 40% of bosses surveyed plan to hire fewer graduates, per the same article. Fortune

The Pulse — reports that Anthropic has dropped the algorithms-style coding interview from its engineering hiring loop, after a hiring manager there concluded it gave zero signal on candidates. If the lab that builds Claude won't test hand-written algorithms, your interview loop is the outlier — and Amazon now writes AI usage into engineer and manager expectations. incident.io also cut AI spend from $4,500 to $1,600 per developer per month between August and now. The Pulse (Pragmatic Engineer)

Design reviews aren't dead — shrink them to two pages. A LeadDev argument that coding agents haven't made design reviews obsolete. In the opening case, a feature went straight from a requirements document to plausible pull requests, and only in code review did anyone notice the approach ignored reliability and operations. The proposed fix: a two-page brief — problem, approach, rejected alternatives, failure modes — required only for changes crossing team or service boundaries, altering stored data or public contracts, or hard to reverse. A ready answer for teams skipping design because the agent planned it: the plan says which files change, not whether the approach is right. Published October 27 by Diego Quiroga of Microsoft. LeadDev

“Code that delivers the requested behavior is not necessarily the right technical solution to the problem.”

Claude Ultra — no longer carries the rolling five-hour Claude Code usage cap — after account verification, only the weekly allowance remains. Anthropic renamed its top consumer plans: Max 5x becomes Max at $100, Max 20x becomes Ultra at $200, each bundling API credits equal to its subscription price. This removes the mid-refactor lockouts your heaviest users complain about, but adds no weekly capacity — Ultra is still ten times Pro's weekly usage. OpenAI raised limits for Pro and Business Premium a day earlier. The New Stack

no-agent-summary — A community gist shows Claude Code re-sends every background agent's full transcript to the model roughly every 30 seconds, by default since August, just to refresh a three-word status line. One team's telemetry put these hidden calls at 16.4% of all Claude Code spend — more than the main conversation itself. On subscriptions this quietly consumes your team's rate-limit headroom; on metered plans, actual money. Independent measurements collected in the gist's comments land between 18% and 26%. GitHub Gist (Flamerix AI)

Flat-rate AI seats are a subsidy — budget for the metered era. Nordhealth CTO James Stanier argues the flat per-seat plans most mid-size teams buy are a subsidy, and the labs are already moving their largest customers to per-token billing — Anthropic's Enterprise plan has had a $20 seat fee plus metered usage since late 2025, and OpenAI moved Codex's enterprise users to token pricing in April. Make consumption visible and match model and effort level to the task now, before the first metered bill lands. AT&T cut advanced-task costs by up to 56% by routing lighter work to open models, for a 2% quality drop. LeadDev

Developers say supervising agents is exhausting. A 137-point Ask HN thread asks whether AI work tires people out, and the replies mark a mood shift: engineers describe fatigue from constant guardrailing, context-switching and reading generated text, with the role now feeling like reviewing output rather than writing it. The strain is supervision load, not output volume — worth reading before your next retro or one-to-one. Commenters openly doubt anyone reviews AI-generated tests, and one says they've shelved personal coding projects altogether. Hacker News

AI-cost fears knock tech stocks for a second day. Intel fell 28% and Alphabet has lost over $500bn; the Nasdaq just closed its worst month since March 2025 as AI-spending fears rattle your vendors. BBC News

OpenAI — now seeks about $20bn at a $1.25tn valuation, down from $30bn at $1.4tn, per FT sources; the WSJ reports roughly $11bn in commitments so far. Techmeme (Financial Times)

Anthropic — may receive only a high junk credit rating after its IPO, Blackstone told investors in its $63bn chip-financing deal — relevant before long-term Claude commitments. ft.com · from the Hacker News discussion

DoorDash — says outside AI agents have produced only several hundred orders, against roughly 3 million daily deliveries — a reality check for any agentic-commerce line in your roadmap. TechCrunch