Gemini 4 Hydrogen lands at $1.50, a team drops peer review, and a GitHub credential heist

🔥The one thing to know

🧭How teams are doing it

• A team drops peer review, and it holds. An engineering lead with 15 years in the job writes that his team stopped peer review a month ago: whoever owns a ticket reviews the PR Claude wrote for it and merges once CI and an automated review tool are clean. A concrete template if review load is your bottleneck — the second pair of eyes becomes the person with the most context, not a teammate paging it in from scratch. He now allows solo work streams too, since an agent can brief him on a colleague's project in minutes; his stated caveat is one month of data. devonestes.com

• Everyone uses AI; barely a third measure it. Two surveys published together: 96% of 305 technology leaders say their companies build software with AI, but only 55% have a clear strategy, and just 31% of 591 engineering leaders measure AI's impact at all. Ammunition if you are pressed to prove value — the industry-wide gap is measurement, not adoption. The share of companies expecting AI to let them hire fewer developers fell from 76% in 2025 to 55% this year, while 62% now expect fewer junior hires. leaddev.com

• Jane Street — The trading firm is trialling new campus-hire onsite rounds with less code-writing and more system design, judging whether existing code is sound, and getting working code out of an LLM. The first big-name public template for what agents change in hiring: judgement about code over typing speed. Most rounds stay unassisted and traditional, and candidates need no polished agent workflow to pass — the firm screens for clear thinking about where the tools help and where they fail. blog.janestreet.com

  “you think clearly about the tools you use: what they're good at, where they fall short, and when to reach for something else”

• Sugabots — The five engineers behind Nitric replaced a single do-everything agent with specialists — Sentry triage, backlog organisation, fix PRs — and open-sourced the shared workspace they built, Sugabots. A workable pattern for cutting night pages while keeping control: engineers challenge each agent's calls, and each narrow role has one clear failure mode to inspect. Alerts used to drag their on-call engineer out of bed three or four nights out of seven; the triage agent runs on Opus 5.5 and may only auto-fix small changes outside areas like billing. thenewstack.io

🛠Agents and plans

• Claude Sheets, Whiteboards and Data — Anthropic added whiteboards, AI-built spreadsheets and dashboards to every paid Claude plan, with exports to Excel, Miro and Asana — a direct push at Microsoft 365 and Google Workspace. Two admin notes: on Team plans all three switch on by default; on Enterprise, Whiteboards and Sheets start off, and Data switches itself on come November 19. Microsoft's stock closed down 2.20% at $533.16 the day after the launch; Anthropic's own product lead admits the spreadsheet features are basic for now and points advanced modelling work back to Excel. claude.com

• OpenShell — Nvidia handed OpenShell, its open-source runtime that runs coding agents inside enforced sandboxes, to the Cloud Native Computing Foundation, the body that stewards Kubernetes. This becomes the neutral home for the sandbox-and-policy layer your platform team may end up running: it uses Cedar, a CNCF policy language, to define which files, networks and credentials an agent may touch. It entered at Sandbox tier, the entry level, with about 380 organisations already contributing; Red Hat, Docker and Canonical were core contributors from the start. techstrong.ai

🧰Tools around the agents

• AgentRun — Parcha open-sourced a control plane over harnesses like Claude Code and Codex, adding durable runs, spend caps, approvals and a learning step that keeps notes and skills for the next run. For teams watching the per-agent token bill: its pitch is that agents re-derive the same knowledge on every run, and you pay for it each time. The team built it after their Greptile review bill jumped from $180 to $6,451 a month; their replacement reviewer costs 18 cents a review and beat Greptile's bug count on 30 unfamiliar PRs. github.com

• DevEx Metrics Compass — Researcher André Meyer mapped more than 120 developer-experience metrics from research and industry into a navigable compass; DX's engineering podcast walks through it in 46 minutes. The antidote if leadership asks for an AI dashboard: choose a small metric set tied to real decisions, retire ones that stop being actionable, and pair telemetry with surveys to learn why numbers moved. One theme for agent-heavy teams: machine-written code is loading more review onto peers, and the prescription is explicit PR expectations plus risk-based review depth. newsletter.getdx.com

💬What people are saying

• Humble Bundle — A Hacker News thread reports that Humble Bundle's rebuilt site dropped the second factor from login, broke ebook downloads and enabled one-click payment, and pins the regression on AI-assisted coding. The argument underneath is the one your team will have too: some say cheap models plus thin engineering oversight did it, others say unskilled hands did — the accountability question is identical either way. Commenters also report login emails arriving an hour late and at least one long-time subscription cancelled on the spot. forums.spacebattles.com · from the Hacker News discussion

  “A computer can never be held accountable, therefore a computer must never make a management decision”

🌍The big AI stories today

• wfx campaign — Stolen logins for 4,837 developer accounts pushed a CI secret-stealer into tens of thousands of repositories; audit branches and rotate everything a pipeline can reach. aikido.dev

• Nvidia — The board is weighing Jensen Huang's exit; about $600bn of market value went this week and CFO Colette Kress leads the rumoured shortlist. trendingtopics.eu

• Anthropic's IPO — Shares priced at $70–$78, valuing it near $1.4tn, with trading possible late this week — your main agent vendor goes public below earlier expectations. theverge.com

💡One takeaway