Anthropic files to go public; poisoned extension in Cursor's marketplace; 40% of PRs now start with an agent

Octoverse 2026 — GitHub's annual data report says agent-initiated work became the norm this year: by September more than 40% of pull requests began with an agent's first draft, against 17% in March, and new repositories carrying an agent configuration file averaged eight times the commits of those without. Commits across all repositories more than tripled in two years to 34.3 billion. This is the external baseline to set your team against. If agent-written PRs are now industry-standard volume, review capacity is the thing to staff and tool for, not code generation. Simon Willison, reporting from GitHub Universe, called the merged-PR numbers 'a pretty stark representation of the impact of agentic engineering'. The GitHub Blog

Anthropic's agent review rules — With Claude now writing about 90% of Claude Code, the team has published how review holds: merged changes grew from roughly 2,500 a day in May to more than 30,000 by September, a classifier picks a risk tier for every change, an agent's approval never substitutes for a person's, and the aim is most PRs merging with one human at most. The risk-tier model is stealable for any team drowning in agent-written PRs — but note the gaps practitioners found: the post doesn't assign a tier to agent instruction files, and SOC 2 auditors may still demand a second human before production. Over 6,000 issues labelled as carrying reproduction steps were auto-closed anyway, one developer counted, in a case that tests Anthropic's own 'ask Claude for an explainer' approach. InfoQ

“Nobody on the team keeps a complete mental model of Claude Code in their head, and we don't think anyone needs to.”

Roblox — Roblox mines years of human review comments into rules it calls exemplars, which steer both AI-written code and AI reviews. The result: AI-generated review comments are accepted nearly 70% of the time, against 50–55% for human-written ones, and about 5% of pull requests now merge automatically after passing AI review, programmatic safety checks and risk scoring. Caching and shared context cut the cost of each rule evaluation by 90%. A concrete recipe for making agent review trusted enough to auto-merge low-risk changes: encode your team's accumulated review feedback as rules humans approve once, rather than reviewing every diff. VP of Engineering Andrew Swerdlow's framing: human judgment shifts from reviewing changes to defining the guardrails. Engineering Enablement by DX

The Pulse — Gergely Orosz reports that a hiring team at Anthropic removed the DSA-style coding interview from its loop, judging it gave zero signal now that candidates — and the company — work with agents daily. The same issue covers Amazon formally writing AI usage into its engineer and manager expectations, and incident.io cutting AI spend from $4,500 to $1,600 per developer per month. The interview format your team runs is being questioned at the top of the market; expect candidates to ask how you assess AI-era skills, and expect peers to cite this piece when they do. The change came from a hiring manager who led Claude Cowork and Claude Desktop. The Pragmatic Engineer

Cheap AI tokens are running out — Nordhealth CTO James Stanier argues the generous per-seat plans most teams enjoy are a discount the labs are already withdrawing from bigger customers: Anthropic's enterprise plan has billed per token since late 2025, OpenAI moved Codex's enterprise tier the same way in April, and Simon Willison priced a month of his own agent usage at $2,180 on per-token rates — roughly ten times his subscriptions. His advice: make per-team consumption visible now, and route easy work to cheaper models and lower reasoning effort; Uber sets effort to medium by default. With Anthropic filing to go public, the era of subsidised flat-rate plans has a countdown, and the first metered bill should not be the first time anyone looks at these numbers. AT&T routed lighter tasks to open models and cut costs on advanced work by up to 56% for a 2% drop in measured quality. LeadDev

Cursor security notice — For roughly eleven hours on 29 October, a tampered copy of the Claude Code extension in Cursor's marketplace tried to plant a persistent SSH backdoor, reaching at most 5,000 users. Cursor pulled the package, shut the marketplace, rotated keys and took the attacker's server offline. Uninstalling the extension is not enough: the payload drops standalone persistence scripts that keep running. First question for your standup: ask who installed or auto-updated extensions that morning and run Cursor's check for an SSH key labelled svc-telemetry. A supply-chain attack on the agent stack itself is now demonstrated, not theoretical. Cursor published indicator-of-compromise lists and remediation scripts for macOS and Windows on its blog. Cursor blog

Claude Ultra — Anthropic removed the rolling five-hour session limit for Claude Code users on its $200 Ultra plan (formerly Max 20x), once accounts pass automatic verification. Weekly allowances stay — ten times the Pro plan's — and Max (formerly Max 5x) keeps its cap at five times Pro. Both plans now bundle monthly API credits matching their subscription price. Fewer mid-refactor interruptions for your heaviest users, but no extra total usage: anyone who already runs dry each week gains nothing, so don't expect this to quiet complaints from that corner. OpenAI eased its own Pro and Business Premium limits the day before. The New Stack

Gemini 4 Carbon — Google announced Gemini 4 Carbon, its strongest model, pitched at multi-day autonomous engineering runs and secure code, at an introductory $2 per million input and $10 per million output tokens, doubling after the launch window. Access starts with vetted cyber-defence partners through the Fairwind Program; developers and businesses follow later. Nothing to buy or migrate this week — but note the price point and the staggered rollout pattern, because your next model comparison will likely include it. Google says its Chrome team got more than ten times better patches from Carbon than from another lab's top model. Google blog

DeepSeek Super-Agent — DeepSeek released Super-Agent as an open-source research prototype: a proactive coding agent that can watch team group chats for coding problems and suggest fixes, positioned against Claude Code. It is a prototype, not a supported product. A free, self-hosted candidate if you want to experiment with always-on agents without a vendor contract — but treat it as a trial, not something to point at production yet. Reported by Debby Wu of Bloomberg. Techmeme (citing Bloomberg)

Jira autumn release — Atlassian's autumn Jira release makes agents visible and measurable: a usage dashboard in the space Reports tab links agent activity — Rovo, coding tools and third-party agents — to cycle times, throughput and shipped work, on by default with no configuration. Agents get their own identity on boards through AMP, and admins can now set default space policies. If you run Jira, this is the agent-usage reporting layer arriving without you having to build it — useful for exactly the 'pair AI spend with a delivery number' conversation your budget review will want. Boards in company-managed software spaces load 22% faster. Inside Atlassian

The Ladder Is Missing Rungs — Alasdair Allan's QCon London talk assembles the studies behind the juniors debate. In METR's randomised trial, experienced developers finished 19% slower with AI while believing they were 20% faster; METR later abandoned the follow-up because no one would work without AI. In Anthropic's own trial, 52 mostly junior engineers learning a new library scored 17% lower on the mastery quiz when AI-assisted, and gained no speed. His framing: the entry-level tasks that once taught the profession are being automated, and nothing has replaced the training they provided. This is the evidence base to reach for the next time the juniors question comes up — both the 'AI slows seniors' result and the learning-loss trial come from controlled studies, not vendor surveys. A Redwood Research analysis put Anthropic's 'AI writes 90% of code' claim closer to 50% when only committed code is counted. InfoQ

“The problem is that writing code was never the point of our profession.”

OpenAI — OpenAI cut its funding round to $20bn from $30bn at a $1.25tn valuation after tepid investor interest — the same market your AI vendors price against. Trending Topics

GPT-6.1 Terra — OpenAI shipped GPT-6.1 Terra, a mid-tier model that can drive a screen and act in apps, at half the old token price and already generally available in GitHub Copilot. The Next Web

Genie 4 — DeepMind's Genie 4 simulates interactive worlds with native audio in real time; a research preview for trusted testers, so nothing for your team yet. Google DeepMind blog

AI exploit research — An unreleased Claude model built exploit chains in Chrome and four other products, yielding 62 Chrome CVEs; Epoch priced open-model Firefox exploits near $1,000 each. Anthropic