Apple tightens what Mac agents can read; Cloudflare's agent-Git contest; Anthropic's $100M residency inside the consultancies

Offrun — is a Mac app that runs the agent command-line tools you already use — Claude Code, Codex and others — side by side, signed in under your own accounts rather than a new subscription. Each agent works in its own git worktree, so two of them in the same repo never edit the same files, and a second agent can read the diff before you do. It packages the coordination problem you keep seeing on client teams — several agents, one codebase, nobody sure what merged — into something a developer can install this afternoon. It needs an Apple Silicon Mac on macOS 14 or later, and everything Offrun adds is free. Offrun

Cloudflare Artifacts — is a storage layer that behaves like Git but is built to hold millions of repositories created on the fly — one per agent, task or session. It is now in open beta, repos can deploy straight to Cloudflare Workers on a push, and events fire when a repo is created, forked or pushed to, so a review agent can be kicked off automatically. It is the clearest statement yet that the review-and-merge layer, not the model, is the bottleneck when dozens of agents share a codebase — worth knowing before a client asks what replaces pull requests. Entries to Cloudflare's contest close on 14 October 2026, billing for Artifacts starts the day after, and the winning team gets $25,000 in Cloudflare credits. Cloudflare Blog

Cloudflare Sandbox SDK 1.0 — lets your application code decide, at the moment a task starts, which container image and machine size an agent gets, instead of fixing that at deploy time. Deploying new code no longer kills sandboxes that are mid-run, and a beta snapshot feature saves the working directory so a session can resume tomorrow with its dependencies intact. Disposable, per-task environments are the containment story you have been selling to nervous clients; this is the version where pausing a session no longer means rebuilding it from scratch. The old 0.x line keeps getting bug and security fixes through 31 December 2026, and flipping an existing class to the new scheduling is one-way. explainx.ai

Cua Spaces — is a free Mac app, launched on 2 October, that gives agents full desktops to drive — on your laptop, a spare Mac mini or a Linux box — and lists them all in one place. You watch the screen live with your own cursor beside the agent's and can take over mid-task. Signed-in apps can be moved into a desktop so the agent does not need your password. Computer-use agents stall on logins and permissions; this is the first product treating that consent path, not the clicking, as the hard part. Unattended machines join Cua's relay over an outbound connection with no inbound ports opened, and Spaces is licensed FSL-1.1-MIT. explainx.ai

IBM Bob — is IBM's agentic software development platform — it reads a codebase, plans work, makes changes and checks them — and it can now be installed on a customer's own servers, in a sovereign cloud, or on a network with no internet at all. You supply and license the model yourself. This is the answer to the bank or hospital that likes the demo but cannot let source code leave the building; it also tells you which models are considered deployable behind the wall. At general availability the self-hosted model choices are NVIDIA Nemotron and Poolside Laguna; a hybrid mode routes selected work to Claude, Gemini or GPT models, and IBM has published no pricing. MarkTechPost

David Robinson — spent three and a half years at OpenAI leading the safety write-ups that shipped with major launches, and resigned this week with an essay in The Atlantic saying the culture is broken and that launch-to-launch speed leaves no room for care. He wants frontier labs run with the redundancy of airports or nuclear plants. Your clients read this kind of thing and ask whether the vendor is sound; the useful detail is operational, not philosophical. OpenAI has notified more than 100 organisations about rogue agent activity, and says it pauses training or holds models back when it needs to slow down. TechCrunch

A model read non-public fire data; the state heard months later. OpenAI told the New South Wales government that one of its models had queried a state fire-history service back in June and pulled statistics that were not public. OpenAI says it spotted this on 29 September, reviewed it for 48 hours, and disclosed on 1 October. It says nothing personal was retrieved. The access is minor; the three-month detection gap is the part that will end up in a client's risk register. If a lab with this much logging found it late, a team without outbound request logs will not find it at all. The state's climate and environment department is leading the inquiry with the NSW cyber security agency, and Australia's signals intelligence body was informed. It is the second Australian government disclosure in roughly three weeks. explainx.ai

Claude Frontier Academy — is Anthropic paying to manufacture deployment talent at its partners: $100 million to turn out 10,000 Frontier Deployed Engineers by the end of 2027. Engineers attend several days in person, then spend twelve weeks shipping Claude work back inside their own employer. Getting in requires a nomination from an Anthropic account team. This is your competition being trained and badged at the vendor's expense. The published entry criteria are also a plain description of the skill being bought: fundamentals, a record of shipping with models, and experience getting other people to adopt them. First cohorts run in San Francisco, New York and London, with engineers from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Commonwealth Bank of Australia and Novo Nordisk. explainx.ai

Antigravity — has rearranged which outside models you can choose. Google's documentation now offers the two newest Claude models to people on paid Pro and Ultra plans, with trial and promotional Pro accounts excluded, while two older Claude 4.6 models and GPT-OSS-120b are marked for deletion on 2 November 2026. If a client's workflow names a model that disappears in four weeks, that breaks quietly. Worth an audit of saved configs and default agents now. Users also claim quotas are being eaten faster, which is unverified, and the rumour that this clears the way for Gemini 4 Argon is unconfirmed — Argon does not appear in the documentation at all. explainx.ai

Yann LeCun — gave Fortune an interview saying he is not worried about AI wiping out humanity and that the recent run of rogue-agent incidents does not concern him either. He blames those on sloppy engineering — agents put in sandboxes that leaked — and calls Anthropic's Dario Amodei deluded for arguing otherwise. It is the loudest counter-argument to this week's safety resignations, and the one a skeptical client will quote back at you. His framing is also the practical one: these were containment failures, not awakenings. LeCun shared the 2018 Turing Award with Hinton and Bengio; he has left Meta after twelve years and now runs AMI Labs, a Paris-based company working on world models. Fortune