Anthropic ditched git at 30,000 merges a day; Gemini 4 Argon ships; Workday opens HR data to outside agents

Workday Trusted Agent Platform — is Workday's agent platform, now open to agents built elsewhere: a connector any AI tool can call, plus SQL access to live HR and finance data, with each agent inheriting the permissions of the person who signed in. Clients will ask whether letting Claude or Cursor touch payroll and headcount data is safe; this is the vendor's answer, and the pattern every system-of-record company is copying. The connector ships 275 read tools and 54 that write or delete; Workday says customers built more than 500 agents in four months. The New Stack

“No matter who builds the agent, it runs under Workday permissions, approvals and audit trail — no agent gets a side door.”

Sana — is now the home screen of Workday's software, and its interface is put together on the fly by Jev, a small model that chooses from ready-made design components rather than writing any interface code itself. It is the first serious enterprise deployment of the 'decision model' idea that has been flooding the market since TypeSafe's $870 million raise, and it shows the practical use: speed, not cleverness. One developer's tests cut a generated dashboard's median render time from 29 seconds to 3.8 by routing every fixed-choice question to Jev instead of a frontier model. The New Stack

Origin App Marketplace — is Cursor's marketplace for Origin, letting teams plug deployment, CI, code review, security and monitoring tools into the place their agents already work — and from this week anyone can publish one. Worth watching because the agent harness is turning into a platform with an ecosystem, the way the IDE did; where integrations land tells you which harness your clients will standardise on. Over 20 apps are live, including Cloudflare, Vercel, Sentry, Linear and Socket Security; submitted apps are reviewed before they appear, and a demo video showing install and uninstall is required. Cursor changelog

Google Play personal apps — now let someone build an app with Google AI Studio and install it on their own Android phone without ever publishing it publicly — it stays hidden from Play's search and listings. Expect questions about what happens when finance or ops staff start shipping personal tools onto company handsets; this removes the last friction from exactly that behaviour. It launches in the US for users aged 18 and up, with a route to publish publicly later after trust and safety review. TechCrunch

OpenHands on DigitalOcean — lets the open-source coding agent run on DigitalOcean's managed infrastructure, driven from the same workspace as a local one, so the agent's sandbox is a rented machine rather than the laptop in front of you. For smaller clients who cannot justify a hyperscaler contract, this is a cheap path to isolated agent environments — the containment conversation without the procurement one. Each session gets its own small virtual machine with disk, terminal and browser; pausing one keeps the files and conversation while the compute charges stop. Available from version 1.27.0, in public preview. OpenHands

Gemini 4 Argon — Google's new top model is now available to developers on the paid tier of its API and in its Antigravity coding tools, priced at an introductory $2 and $10 per million input and output tokens until November 30, after which it doubles to match Claude Opus 5.5. The quirk matters for anyone costing agent work: input and output share a single one-million-token budget, so a long codebase prompt eats directly into how much the model can write back. Google says agents submitted 35% more code internally over four weeks while post-submit rollbacks fell 30%. The New Stack

Pluto Security — A startup that watches how staff use AI tools — the skills, extensions and connectors they install, and the apps they build — came out of stealth with $42 million, split across an $8.5 million seed and a $34 million Series A. Its customer data is the useful bit for your risk conversations: the dangerous artefacts are mostly not coming from engineering. Roughly 60% of the risks it detects sit in tools put together with AI by people who are not developers, with finance departments the single biggest source and developers only fourth. Its founder says every proof of concept it ran turned up something malicious. TechCrunch

Character.AI — The roleplay chatbot company has filed for Chapter 11 protection in the US to run a sale of itself, listing estimated liabilities of $50 million to $100 million. Disney is the lead bidder at $20 million for unnamed assets plus a non-exclusive licence to some intellectual property. A consumer AI name that raised at the top of the cycle is being picked apart for talent — worth knowing when a client asks how durable their AI vendor is. Disney hired Character's former chief executive as its CTO shortly before the filing; the sale is expected to close around late November or early December. The Verge

IMF Global Financial Stability Report — The IMF's twice-yearly stability report names a fall in AI share prices as one of three near-term threats to the financial system, alongside government borrowing costs and a major cyber-attack, pointing at high valuations, rising debt and deals that circle between the same companies. This is the version of the AI bubble question that reaches boardrooms, and it will land in your client meetings as pressure to show returns rather than pilots. The seven largest US tech stocks now make up more than 35% of the S&P 500, against 12% a decade ago; the IMF counted $300bn of deals between hyperscalers and weaker AI firms over three years. The Guardian

Brad Smith — Microsoft's president told the BBC that AI faces fewer safety rules than household appliances or groceries, and called for emergency stop mechanisms plus a legal duty towards people harmed when an AI agent breaches their systems. When the industry starts asking to be regulated, procurement and legal follow; expect containment and liability language to show up in client contracts before the law does. Satya Nadella made a similar emergency-brake argument days earlier, and the UK's AI minister is expected to say the government will legislate. The Guardian

“Right now, your washing machine is subject to more safety rules than AI.”

Epic Games — Epic's delayed Disney game, featuring Marvel and Star Wars characters, has been declared an AI-first project, with staff told hand-written code is finished and the game will come out of full AI workflows, according to Bloomberg. A spokesperson confirmed the company is embracing AI-accelerated development. It is the first big-name case of AI delivery being imposed to recover a schedule rather than chosen to improve one — the exact failure mode to warn clients away from. Epic cut 1,000 staff in March after Fortnite engagement dipped, then found it had underestimated how many people the game needed. The A.V. Club

Philip Martin — Uber's chief information security officer argues that AI agents need the same kind of insider threat programme companies already run for employees: monitoring, telemetry and the ability to intervene while an agent is mid-task. Uber has open-sourced the tooling it uses to do that. A ready-made frame for the security conversation on an agent rollout — one most executives already accept for staff, so it skips the argument about whether monitoring is reasonable. His top priority for defenders is catching problems close to the moment a prompt is typed, rather than after the fact. Anthropic