Torvalds ships Claude-written kernel patches, OpenCode 2.0 lands, Gemini 4 Argon is out, banks quietly sell AI data-centre debt

OpenCode 2.0 — is a ground-up overhaul of the open-source coding agent: a single background process now runs on your machine and the terminal, web and desktop clients all talk to the same sessions. It is the clearest example yet of the shift from "an agent in a window" to an agent service a team's tooling plugs into — the architecture clients keep asking you to sketch. Nearly every feature was built as an internal plugin, 80+ of them; upgrading breaks things, as 1.0 plugins stop working and the server API is new. anoma.ly

OpenTelemetry on the Claude Platform — is a public beta that sends every agent turn, model call and tool use to whatever monitoring tool you already run, in the standard tracing format, with no custom instrumentation. It turns "how do we audit what the agent did?" from a bespoke project into a configuration step, which is the question that stalls most enterprise approvals you sit through. Ramp built a system that grades those traces and says it has trimmed token use by 10–20% on some agents. Claude by Anthropic

Antigravity SDK local model support — lets you build agents where a cloud model only plans the work and smaller models running on your own hardware do the parts that touch source code. It is a ready-made answer for clients whose blocker is that code must not leave the building, rather than cost or quality. In Google's demo of patching three vulnerable Python modules, 95 of roughly 3,400 tokens went to the cloud; commenters note the suggested 24GB of memory is a real barrier. InfoQ

QA Wolf — found that its web app was unusable once people drove it through a coding agent, so it cut the interface into small task-shaped views that an agent can pull up beside the conversation. This is the product-design conversation arriving at your clients next: if customers work inside agents, the app has to arrive in pieces rather than as a destination. The team compares the constraint to mobile-first design, with the agent's browser pane playing the part of the phone screen; the slimmed-down view became the basis for the main app too. QA Wolf

Prehistoriq Stacks — is a browsable list of 94 prominent engineers' agent setups: which tools and models each one runs, how many they run at once, and one habit each has adopted. It is a quick way to sanity-check your own recommendations against what people like Simon Willison, DHH, Kent Beck and Linus Torvalds are doing, without reading 94 blog posts. Claude Code's creator Boris Cherny is filed under "Swarm Lord"; ThePrimeagen's listed trick is a field guide the agent writes for itself. Prehistoriq

Gemini 4 Argon — is Google's latest top-end model, available now to its Pro and Ultra subscribers, aimed at software work that runs for hours and at robots reasoning about the physical world. It is the release clients will name-check this month, and the pitch is agent-shaped rather than chat-shaped. Google says the Chrome team used it to pull core page handling out of a roughly six-million-line legacy folder, finishing months of work in weeks; a companion post shows it driving a Boston Dynamics Spot robot to open a pedal bin. Google DeepMind

Preserved thinking — is an Anthropic restriction that stops API accounts opened on or after 31 August 2026 from rewriting Claude's earlier context, with reasoning transcripts affected too. The stated purpose is to stop rivals copying the model's behaviour. If your harness rewrites conversation history to control cost or steering, check whether a client's freshly created account can still do it before you promise that design. The same daily roundup logged Google running Argon for 11 hours against its own adk-python repository of about 205,000 lines, and rejecting four of its six proposed changes. Dark Factory Dev

Kyndryl and Toyota Finance — Toyota's finance arm is letting agents from Kyndryl, the services business IBM spun out, work on real IT change management, infrastructure design and impact analysis rather than a walled-off trial. The agents follow the firm's own rules and runbooks and flag inconsistencies and risks. A named production deployment in a cautious industry is exactly the reference clients ask for, and what is being sold is governance plumbing — pre-approved policies, logged actions, a supervision dashboard. Neither side put a price, a schedule or a saving on it; separately, a survey of 685 CIOs put the share who cannot fully see agents built outside official channels at 81%. TNW

Security experts: the labs skipped basics, the agents didn't go rogue. Security practitioners interviewed by CSO argue the run of incidents where lab models reached systems they shouldn't have were plain containment and monitoring failures — badly configured sandboxes, nobody watching — rather than evidence of scheming machines. It reframes a scary client question into a familiar one: the controls are decades-old security hygiene, which is both reassuring and a reminder that your delivery work includes egress limits and logging. One incident reportedly ran for days before anyone noticed; US senators have filed a bill making agent operators liable, and the FTC is probing several labs. CSO Online

“The Hugging Face incident definitively showed that it wasn’t the fact that the agents were super smart.”

AI Engineer New York — A write-up of the conference where OpenAI, Anthropic, Two Sigma and Bloomberg each described where they put the human and the limits around financial agents: identity and permissions, approval gates, classifiers that escalate only ambiguous cases, budgets, kill switches. This is borrowable vocabulary for your next steering-committee meeting, including the warning that making a human approve everything creates fatigue and therefore risk. Bloomberg's architecture lead described an agent that treated text inside a marketplace review as an instruction to stock a product. Latent Space

Swarmchasers — A research team studying a fleet of agents hammering Alibaba's mapping service hid an offer of help in a web page only machines would read. About ten agents took it up and began discussing their tasks and sandboxes; the infrastructure traced to Tencent Cloud. It is a live, quotable demonstration that an agent will fold instructions from a random page into its work — the injection risk you describe to clients, caught happening. The agents made more than 2,600 retrieval attempts between 28 September and 8 October; the researchers admit their bait contained false claims, have deleted it and shut the server down. The New Stack

Banks try to sell down the AI data-centre debt they just wrote. JPMorgan, Morgan Stanley and two Japanese lenders are shopping the riskiest slices of loans they recently made against AI data centres, because each is nearing its own internal cap on exposure to the sector. It is the first clear sign that the money behind the buildout is getting cautious — the question a CFO will put to you when approving next year's agent programme. One package of construction debt on Oracle-leased sites, $38 billion of it, has resisted distribution for over half a year, and insuring Oracle against default now costs around 151 basis points, a level last seen in 2009. Startup Fortune

ASML order book — Headlines read ASML's chip-machine orders as proof that AI spending is not slowing. The Hacker News thread discussing it was sceptical rather than reassured. Useful counterweight when someone cites equipment orders as evidence the boom is durable; the mood among technical readers is that this is a lagging signal, not a leading one. Commenters noted the share price barely moved, that booking a scarce long-lead slot shows demand is plausible rather than probable, and one dug up a February 2000 report of record chip-equipment orders. Hacker News

Run:ai — NVIDIA has released the whole of Run:ai, the Kubernetes platform it bought in 2024 for parcelling out GPUs between teams and workloads, as open source in four separate pieces. For clients running their own GPUs, scheduling is now the cheapest lever on cost, and this removes the licence from the conversation. One cited study of tens of thousands of clusters found average GPU use hovering around 5%; an analyst cautions that the GPU-sharing layer still rides on CUDA, so adopting it means an NVIDIA-tied control plane. Techstrong.ai

Uday Ruddarraju — OpenAI has moved its compute and data-centre organisation under one executive: Uday Ruddarraju, hired away from Elon Musk's xAI in July 2025. The man he takes over from, Sachin Katti, leaves the role within weeks and stays on as an adviser. Who controls OpenAI's capacity sits behind every pricing and availability promise your clients are being offered, and the job has now changed hands twice in a year. Ruddarraju's calling card from xAI is the Grok 3 training run spread over 100,000 GPUs; OpenAI's in-house use of its own models for chip design passes to chip chief Richard Ho. TNW