Gemini 4 Argon at $2/$10 (but you can't buy it yet), Shopify Canvas, Pi 1.0, and Mollick admits he was wrong about managing agents

Shopify Canvas — lets a merchant build their store by talking to Shopify's assistant Sidekick, with the result rendered as the store's actual code as it changes, not a mock-up — so animations and different screen sizes behave for real. This is what agentic building looks like once it stops being a developer tool and becomes a default inside the software your clients already pay for. The first release skips third-party themes, app blocks and extensions, markets, translations and theme updates, and works on desktop only. TechCrunch

Pi 1.0 — is a deliberately small, open coding agent you run in your terminal, now declared stable after months of hardening, plus an experimental sibling for agents that need to keep running over hours rather than one sitting. The interesting part is the posture: the team says it waits until a feature has proven itself before adopting it, which is the opposite of the weekly churn that makes agent tooling hard to standardise on for a client. Hundreds of thousands of people use it weekly; both Pi 1.0 and the long-running package, Pi Durable, are MIT licensed. Earendil

aweb — gives agents a permanent address, a mailbox that survives their session ending, and a wake-up signal when something arrives — so an agent on your laptop can hand work to one on a colleague's machine and get an answer later. Most multi-agent setups today only work inside one process on one box. This is the plumbing question clients hit the moment two teams want their agents to cooperate without sharing a login. Addresses look like domain/name, with trust anchored in DNS and signed membership certificates; wake-up integrations exist for Claude Code and Pi, and you can self-host the server with Docker Compose. aweb

Strands Decider 2B — is a small open-source model from Amazon that doesn't write prose at all: it picks between options you've already defined and tells you how confident it is in the pick. It's the cheap answer to a question you get constantly — do we really need a frontier model for every step of a workflow? For routing decisions, a local two-billion-parameter model with a confidence score is faster and less expensive. It was a side project by AWS engineer Marc Brooker after he saw TypeSafe's Jev, it runs locally, and dozens of copycat decision models have appeared since Jev launched. TechCrunch

“I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart.”

Armadin — raised $255.5 million at a valuation above $2.5 billion for a security startup that swaps the scheduled penetration test for always-on swarms of agents that chain weaknesses together to break in. Your security-conscious clients are about to be pitched this. It's also the clearest sign that "agents attacking you continuously" is now a funded product category, not a thought experiment. The Series B, led by Andreessen Horowitz and Accel, landed just six months after a $190 million Series A in March. Founder Kevin Mandia sold his previous company, Mandiant, to Google in 2022. TechCrunch

GPT-Synopsys — is a jointly developed model from OpenAI and chip-design software maker Synopsys, trained to operate Synopsys's design tools directly rather than just advise an engineer who operates them. This is the template for vertical agent deals: the model vendor licenses the incumbent's tooling, the incumbent keeps the customer, and both share the revenue. Expect the same shape in other regulated, tool-heavy industries. The multi-year agreement includes revenue sharing and joint go-to-market; the model runs on OpenAI-hosted infrastructure, and Synopsys says customer design data is not used for training. Synopsys

OpenAI dismisses three safety researchers over leaks. OpenAI parted ways with three researchers on its safety team who, according to the Wall Street Journal, shared confidential information with an outside AI safety organisation. The company says an investigation found they mishandled sensitive information; none of the people or the organisation have been named. It feeds the question clients are already asking about whether the fastest-moving lab is also the one you want handling your data — worth knowing the sequence rather than the headline. The dismissals came two days after the New York Times reported executives had brushed aside internal safety warnings, and in the same week OpenAI scrapped the planned GPT-6.1 Astra launch. TechCrunch

Six labs sign a voluntary safety checklist at the White House. Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed a joint commitment on 30 September, with Trump counter-signing, promising four layers of oversight: internal controls, a team that owns them, an independent external assessment, and a board committee that reviews both. It is voluntary and carries no penalties. It has no legal teeth, but those four layers will be pasted into enterprise security questionnaires long before any law exists — so you'll be asked which of your upstream vendors signed. Microsoft, Amazon and Apple are absent from the signature page, which also carries a typo: "President of the Unites States". explainx.ai

Claude for Government — is now generally available to federal and state agencies after a public beta since July, running in a FedRAMP High authorized environment, with the Claude Code command line and the Microsoft 365 integration in early access alongside it. Public-sector clients can now stop treating agentic coding as a pilot with a compliance asterisk. It also raises the bar for what commercial buyers in regulated sectors will expect as standard. Agencies pay for what they use rather than per seat, against a ceiling they set in advance; admin actions land in an audit log, and sensitive operations on Anthropic's side require sign-off from two people. Anthropic

GPT-5.5 is being switched off in ChatGPT and Codex. OpenAI's changelog says GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on 14 October 2026, across consumer, Business, Enterprise and Edu plans, with replacement models named. A small date with real consequences: if a client's workflows or internal docs pin that model by name, somebody needs to go look this fortnight rather than after the fact. The retirement does not apply to the OpenAI API, so API-based integrations are unaffected for now. Dark Factory Dev

Brian Chesky — argues in a TechCrunch interview that a chat box is the wrong interface for shopping — too few options, too many turns — and that the real missing piece is an operating system built for agents, so apps and agents can interoperate without one-off business deals. A useful counterweight when a client assumes the future is a single assistant doing everything. He also says Airbnb bookings through today's consumer agents work poorly. Airbnb will spend the next three to six months exploring "multiplayer" interfaces a group can use together, and plans to roll out voice agents this fall. TechCrunch