Apple rations its engineers' AI budgets, Meta swallows Conductor, and a $100/month agent fleet

🔥The one thing to know

🧭How teams are doing it

• GitHub — Pull requests, git operations, issues, the API and Actions all threw elevated errors on November 4, from roughly 09:49 UTC until GitHub declared the incident resolved at 11:50 UTC. Agent pipelines lean on the API and Actions, so a short GitHub blip now stalls a whole team's automated delivery; if yesterday looked broken, it was GitHub, not your setup. GitHub says a root-cause analysis will follow. GitHub Status

• 17 Claude Code bots — The operations lead at a Korean B2B distributor has published the full bill for seventeen bots doing daily work: a $100 Claude Max subscription, no API spend, and an ordinary office MacBook kept awake — no servers, no containers. The budget lesson is that most internal "AI bots" need no model at all — only three of the seventeen call one — and the real cost is human: about an hour a week writing their rules, down from ten hours early on. The three model-using bots shell out to the Claude Code CLI rather than calling the paid API, so they run on the same subscription. Bots in Production

  “US$100 a month, one laptop that doesn't sleep, and about an hour a week of my time. The bill was never the hard part.”

• Gemini chat history — A widely discussed Hacker News post spells out that on personal Google AI subscriptions, saved Gemini chats feed training unless you switch chat history off entirely, and some transcripts are read by human reviewers and kept up to three years. These are personal accounts, not workspace ones, but staff mix them; if your policy tolerates any Google AI plan, the rule to repeat is that confidential material never goes into a personal Gemini chat. The thread drew 606 points and 153 comments in half a day. Hacker News

• LLMJacking — Google's threat team and Mandiant describe a mature criminal trade in hijacked enterprise AI access: stolen cloud tokens and API keys, resold through underground routing services at a fraction of market rates. Your AI spend is now an attack target — a leaked key doesn't just expose data, it bills you. Model keys belong in secret managers, not repos, and billing alerts should fire long before the invoice does. One victim absorbed nearly $1 million in fraudulent calls, while the stolen-access service retails its unlimited plan for about $60 a month. Google Cloud Blog

🛠Agents and plans

• Conductor — Meta has bought Conductor, the app for running several coding agents side by side. Its cloud is winding down: read-only on November 13, data deletion from November 20, and the local app gets fixes only until early 2027. If anyone on your team relies on the cloud version, that migration is due this fortnight — and it is a reminder that a small agent vendor can be bought out from under you mid-project. Active cloud subscribers get a refund; the agents Conductor deploys were never in-house, running instead on Anthropic and OpenAI models. Conductor

• Cat Wu — Cat Wu, who ran product for Claude Code at Anthropic until last week, says she is joining OpenAI to lead product management for its core products; she announced the move on X. The person who shaped the agent tool many of your developers use daily now shapes its biggest rival's roadmap — worth weighing when you decide which tool to standardise on beyond this year. Business Insider reported the move; Wu was one of Anthropic's most visible product managers. Business Insider via Techmeme

• Replit Pro — Replit's new Pro agreement, applied to existing customers on October 9, added mandatory arbitration and a class-action waiver; the thirty-day opt-out window closes on November 8. Missing it means giving up the right to take a dispute with your agent-tools vendor to court, and it is a five-minute job per seat while the window is open. Opt-outs go by email, one per person — the agreement forbids opting out on a colleague's behalf. Read the Diff

• Claude auto-connect — A beta for Claude Team and Enterprise customers lets admins approve a plugin once and have every employee connected automatically: Claude issues the signed identity proof that participating services check, so people stop re-authorising tools one by one. It removes the login chore that slows agent rollouts, but it adds a second place to revoke access — deleting a leaver from your identity provider alone no longer cuts their Claude-issued entry to connected services. Fourteen services signed up at launch, including Datadog, Linear, Notion and Supabase. The New Stack

🧰Tools around the agents

• Honeycomb Innovation Week — Honeycomb's write-up of its Innovation Week talks frames agent management as a management job: set the outcome, widen what each agent may do only as its results earn trust, and monitor the fleet rather than chasing individual runs. About two thirds of teams in Honeycomb's survey have already seen agents cause a production incident, so fleet-level cost and error monitoring has stopped being optional. The failures don't trip alerts: two agents can hand a refund back and forth forever while every error metric stays green. Honeycomb

  “Engineers are still on the hook, legally, morally, ethically. We own our work.”

💬What people are saying

• Agile in the AI era — A LeadDev essay by Alan Sharp, an engineering manager at Arbor Education, argues that as AI compresses coding, the squeeze shifts to requirements, testing and customer collaboration — so Agile's values matter more even as its ceremonies like sprints and estimation fade. If your review queue and test stage are the new constraint, that is where management attention belongs: crisper tickets, faster reviews, and someone actually exercising finished features end to end before release. On his teams, work that once took weeks now reaches code review within a day using agents like Claude Code and Cursor. LeadDev

  “AI-assisted development is moving the bottleneck from the coding stage to somewhere else in the SDLC.”

🌍The big AI stories today

• Marvell / Lumentum — Marvell is buying optical-networking firm Lumentum in a $115 billion all-stock deal, betting AI's next bottleneck is moving data between chips and data centres; closing is expected in the first half of 2028. TechCrunch

• US midterms — At least 33 local measures restricting AI data-centre build-outs won at the midterms, with more still leading — the capacity crunch your model vendors face just picked up a political edge. Tom's Hardware

• Gemini 4 Carbon — Google unveiled Gemini 4 Carbon and is releasing it first to vetted cybersecurity defenders because of its strength at exploiting software flaws — capability-gated releases are a new pattern. Berkeley RDI

💡One takeaway