A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 19 STORIES

AI news, 14 November 2026

From the community

Google Antigravity turns multi-agent jobs into scripts you can review and re-run

Google is rolling out Dynamic Workflows in its Antigravity coding tool. You type /run-workflow and the agent writes a short Python script that hands work between several sub-agents, some running side by side and some one after another. You can read the script before it runs and save it to run the same job again later.

Why it mattersLong agent jobs often wander because each step depends on how the AI reads the last prompt. Writing the plan down as code makes the steps predictable and repeatable, which matters if you want a team to rely on the same research, build or reporting routine every week.

Sourcesx.com post ↗

Gemini can now read your Google Wallet to answer money questions

Google added Google Wallet to the apps that Gemini's Personal Intelligence can connect to. Once linked, and optionally with your bank and card accounts added through Plaid, Gemini can chart your cash flow and budgets, find forgotten subscriptions and price rises, and suggest a holiday budget based on last year's spending. Google says Gemini can only read the accounts, not move or invest money.

Why it mattersPersonal finance is becoming a standard assistant feature, following OpenAI's finance tools in ChatGPT. For small business owners and freelancers it is a quick way to spot wasted subscriptions, but it means granting an AI read access to financial data, so check the privacy terms first.

Sourcesx.com post ↗

Google Cloud adds new AI chips and opens the system it used to train Gemini

Google Cloud announced two new chips in its eighth TPU generation: TPU 8tx for secure, confidential AI use and TPU 8i Max for very fast answers over very long inputs. It also made Cloud Pathways, the DeepMind software that ran Gemini's training, generally available to customers, and added tools to split and combine chip capacity on the fly and to move model data faster. Performance descriptions are Google's own.

Why it mattersThis is infrastructure news, mainly for companies training or running their own large models on Google Cloud. For most businesses it signals more choice and possibly lower costs for heavy AI workloads, and a confidential-computing option for sensitive data.

Sourcesx.com post ↗

Sierra releases Poppy 1.0 so personal agents can work with companies

Sierra announced version 1.0 of Personal Agent Protocol, or Poppy, on 13 November 2026. Sierra describes it as an open standard for how a personal AI agent, acting for one person, finds a company, says who it is, and learns what it is allowed to do there, through the company's website, APIs or the company's own agent. Sierra says the protocol is largely built on existing standards, including MCP, OpenAPI, UCP and OAuth, and that companies can adopt it in stages, starting with the basics.

Sierra says Shopify has enabled Poppy for every merchant on its platform, and that every Shopify store now automatically publishes a Poppy file, so merchants do not need to do extra implementation. In Sierra's account, Poppy handles the moment an agent arrives and identifies itself, while Shopify's existing Universal Commerce Protocol handles carts and checkout. Sierra also says Shopify co-authored Poppy's commerce extension, which links the two.

Sierra says Poppy is now on for everyone who uses Muse, Meta's AI agent, which "automatically looks for Poppy first on every site it visits". It also says 60 new companies have joined as design partners, bringing the total to more than 100, and that several design partners are now going live, including Mercury, Synchrony and Turkish Airlines. The same post describes Mercury and Turkish Airlines in terms of live uses, such as checking balances or fares; the Synchrony section is more tentative, and talks about work that "explores" how a personal agent could consider financing offers. Sierra's list of new design partners includes Manus, whose account later shared Sierra's announcement.

Sierra says any company can implement Poppy on its own, and points to an implementation kit with a validator. The official specification, though, is still labelled draft 1.0. It starts from version 0.4, the 1.0 release candidate, and says fixes for problems implementers find can include changes that are not backward compatible. The Next Web, reporting the same announcement, also notes that the rules are not final.

We have not independently tested Shopify merchant files, Muse's behaviour on live sites, or the deployments Sierra names. We have not previously covered Poppy.

Last checked: 14 November 2026.

Why it mattersPersonal agents are already arriving at company websites, and until now each site has had to guess who they are and what they may do. A shared discovery file is a practical first step, especially if Muse really does look for one first. The specification is still a draft, though, so treat today's implementations as early, and possibly short-lived if the protocol changes.

Sourcessierra.ai ↗ · personalagentprotocol.org ↗ · x.com post ↗ · x.com post ↗ · thenextweb.com ↗

Claude Code's project coordinator now reports on parallel threads in chat

Anthropic's developer account, @ClaudeDevs, announced updates to Projects in Claude Code on 13 November 2026. In a project, Claude coordinates work across parallel threads. Anthropic says the coordinator is now much more proactive: it answers questions in the project chat and posts status updates on ongoing threads.

The same thread announced four other changes. /btw and /fork now work inside projects, so a side question can get its own thread and a new thread can be started from the current one's state. A reminder can be set on a thread, or asked of Claude; the thread is then moved aside until that time, when it returns and is mentioned in the project chat. An individual thread can be shared as a read-only link with anyone signed in to Claude, from the thread menu; threads that had access to your computer cannot be shared. The threads panel now has a row of recently opened, updated or pinned artifacts, and optional thumbnails on each thread.

@ClaudeDevs said the update is on web and desktop, and in the latest mobile app, and that Projects is in public beta on Pro and Max, with a waitlist at claude.com. Anthropic's official Projects documentation, checked on 14 November 2026, describes Projects as a public beta on Pro, Max and Ultra plans, rolling out gradually, and not yet offered on Team or Enterprise. The same documentation still says a project belongs to one user and that thread transcripts do not have the share option other cloud sessions have — which conflicts with the official share announcement. We have not independently tested the new coordinator, share or reminder behaviour.

Claude Code's changelog for version 2.1.325, also dated 13 November 2026, does not list these Projects changes; that entry covers the CLI, VS Code and related releases.

Last checked: 14 November 2026.

Why it mattersThe point of Projects is that Claude, not you, keeps track of several pieces of work at once. A coordinator that reports in the chat is useful only if those updates are timely and accurate, and if you can still find the thread that needs you. The official docs and the announcement do not yet agree about sharing, so treat share-link behaviour as unverified until the documentation is updated or someone tests it.

Sourcesx.com post ↗ · x.com post ↗ · code.claude.com ↗ · claude.com ↗ · code.claude.com ↗

Google Stitch tests a five-second product-flow experiment

Google's Stitch account announced a rapid design experiment on 11 November 2026. It says a rough prompt or screenshot becomes a medium-fidelity, multi-screen product flow in about five seconds, that a tap produces new variations in about two seconds, and that one click sends chosen screens into the full Stitch tool for deeper UI polish. Google's blog the same day, introducing the Gemini Decisions API (powered by Gemini Decision 1), calls this the "rapid canvas experiment in Stitch" and, in a caption, "Stitch Rapid". It says the experiment uses the Decisions API to generate playable wireframes and to apply edits made by voice or tap almost instantly.

Stitch describes this as an early MVP, and says favourite parts will later be folded into the full Stitch experience. A follow-up post points people to a banner inside Stitch, or to stitch.withgoogle.com (which redirects to stitch.google.com). It also says the rapid scratchpad is not saved between sessions, so work should be downloaded or sent to Stitch before the tab is closed. Opening that URL led to a Google Account sign-in. Stitch itself is marked Beta on the official site, and its official account says the product can be tried free of charge; we have not separately confirmed pricing or regional limits for the experiment.

The five-second and two-second figures, and the claim that a first flow can span seven design systems, are company-reported. We have not timed them or tested the handoff into full Stitch.

We already covered the Gemini Decisions API launch, including its price and limits, on 11 November: TypeSafe's Jev came first. OpenAI and Google have followed with decision APIs. This item is only about the Stitch experiment.

Last checked: 14 November 2026.

Why it mattersEarly UI work is usually slow enough that you switch tabs while you wait. If a first flow really arrives in a few seconds, the useful test is whether the screens hang together and survive a handoff into full Stitch — not the stopwatch claim on its own. The experiment is temporary, and sessions are not saved.

Sourcesx.com post ↗ · x.com post ↗ · blog.google ↗ · stitch.withgoogle.com ↗

ChatGPT Wallet lets OpenAI's dots agents pay for things, with your approval each time

OpenAI launched ChatGPT Wallet on 13 November, built with American Express, Mastercard and Visa. You add a card once; when your dot (ChatGPT's always-on agent) is ready to check out, you see the merchant, items, total and card, then approve with Face ID or a fingerprint. The agent and the shop get a one-off payment token rather than your real card number, and your card's usual rewards and dispute protections still apply. It starts with ChatGPT Pro subscribers in the US aged 18 and over; OpenAI says spending limits and pre-approved categories will come later, and it's also publishing guidance for businesses so their checkouts don't block legitimate agent purchases. Shared by OpenAI's @TheRealAdamG.

Why it mattersIf you sell online, some of your customers will soon send an agent to buy for them. Check that your checkout doesn't treat every automated visit as fraud, and keep product, price and stock details clear on the page so an agent can read them.

Sourcesx.com post ↗ · openai.com ↗

DeepSeek opens its app's search, file reading and voice features through its API

DeepSeek made the features behind its own chat app available to developers on 13 November: web search with citations, a deeper multi-step research mode, reading uploaded files and links, an optional content-safety check, conversation titles, and speech in and out (plus standalone OCR, speech-to-text and text-to-speech). They work through three common API formats, including Anthropic- and OpenAI-style endpoints, using the deepseek-flash model. DeepSeek also published an open-source demo chat app that can import your DeepSeek chat history.

Why it mattersBuilding a branded chat assistant with web search and document reading used to mean stitching several services together; this bundles them in one low-cost API. If you go this way, check where your data is processed and whether that suits your clients.

Sourcesx.com post ↗ · api-docs.deepseek.com ↗

Perplexity's Teammates, agents that own ongoing work, arrive on the Mac

Perplexity brought Teammates to its Mac app on 13 November for individual Max subscribers ($200 a month). Instead of answering one question, a Teammate takes on a responsibility, such as triaging bug reports, answering support email or tracking competitors, then plans the tasks, brings in other agents when needed and comes back for your review. Each runs on its own isolated computer and uses the tools you've connected. It's still a research preview; usage is free during a promotional period and then comes out of Max credits.

Why it mattersThis is the same shift as ChatGPT's dots: from asking AI questions to handing it a job. Start with one low-risk, repetitive responsibility and a clear rule for when it must check with you, like "ask me before any refund".

Sourcesx.com post ↗ · perplexity.ai ↗

OpenAI launches Daybreak, security agents that test web apps and draft fixes

At its Cyber at the Frontier event on 12 November, OpenAI announced two security products. Daybreak Red Team uses a team of agents to probe your web applications for weaknesses, confirm reported bugs and check whether a fix actually stops the attack; it's in early access. Daybreak Defense Factory traces vulnerabilities from your security tools through your code and prepares patches as draft pull requests for engineers to review, nothing is deployed automatically; it's a preview plugin in ChatGPT for workspaces with Daybreak access. Shared by OpenAI's @thsottiaux.

Why it mattersAutomated pen-testing has been out of reach for most small teams. If you run a customer-facing web app, it's worth asking your developer or IT partner whether a tool like this could replace or supplement an annual manual test.

Sourcesx.com post ↗ · openai.com ↗

ChatGPT dots start reaching Europe, officially for business plans

@kimmonismus and @Mr_Salio reported on 13 November that ChatGPT's dots agents are appearing for Pro subscribers in Europe, despite a banner saying they aren't included. OpenAI's help page still lists personal Pro access as excluded in the EU, Switzerland and the UK, while Business Premium and Enterprise workspaces in supported regions can use them. So treat personal access in Europe as a gradual, unofficial rollout for now.

Why it mattersIf your team is in Europe and wants to try dots for work, the official route is a Business Premium or Enterprise workspace, where your admin controls access. Enterprise billing for dots starts on 18 November.

Sourcesx.com post ↗ · help.openai.com ↗

Thinking Machines explains how 4-bit training sped up its new model

Mira Murati's Thinking Machines Lab published how it trained Inkling 2 Small, the open-weights model it released this week, using a native 4-bit forward pass on Nvidia GB300 chips from the first step of pretraining to the last round of reinforcement learning. Murati says this was one of several changes that took pretraining from weeks to 68 hours; that's the company's own figure.

Why it mattersCheaper, faster training is a big reason capable open models keep getting less expensive to use. You don't need to act on this, but expect more strong models you can run through low-cost providers.

Sourcesx.com post ↗ · thinkingmachines.ai ↗

How OpenAI keeps hundreds of always-on agents safe inside the company

OpenAI published the controls it uses to run hundreds of persistent OpenClaw agents internally, more than 500 doing background work such as opening code changes from product feedback. Its rules: keep the part that configures agents separate from the agents themselves, sandbox what each agent can reach on the network and on disk, keep passwords and keys in a broker the agent can't see, have a second AI review agent actions, give each agent its own identity, and log everything. Shared by @steipete.

Why it mattersTreat an always-on agent like a contractor: give it its own account rather than yours, only the access the job needs, and a record of what it did. That checklist works whether you use OpenClaw, dots or another agent.

Sourcesx.com post ↗ · developers.openai.com ↗

Exa's AI search now runs inside the EU

Exa, a search engine and API that many AI agents use, now serves its Search and Contents APIs from Stockholm, with queries, results and page contents stored and processed in the EU. EU accounts are locked to the EU endpoint, with no silent fallback to the US, and requests aren't retained by default. It's for enterprise customers for now, and Exa says its benchmark scores match the US service. @kimmonismus called data sovereignty one of the most underrated bottlenecks for AI in Europe.

Why it mattersIf GDPR or a client's security team has kept you from using US-hosted AI search, this removes one blocker. Note that account metadata may still be kept outside the EU.

Sourcesx.com post ↗ · exa.ai ↗

Vercel's Workflow SDK 5 is built for fast agent loops

Vercel released version 5 of its open-source Workflow SDK, which keeps long-running jobs and agent loops going through crashes, deployments and long waits. Every payload is now encrypted by default, runs can live in a chosen region (for example Frankfurt) with failover, and Vercel's own benchmarks show 1.6 to 25 times less overhead than version 4. A Python version is in beta.

Why it mattersIf a developer builds agents or automations for you, durable workflows mean a crash halfway through doesn't lose the job or double-charge a customer. Region choice also helps with data rules.

Sourcesx.com post ↗ · vercel.com ↗

agent-browser gets smaller and starts almost instantly

Vercel's agent-browser, a command-line browser that coding agents use to open pages, click and fill forms, reached version 0.43 on 13 November: up to 7.5 times smaller, starting in about 2 milliseconds, and installed straight as a native program. Shared by its maker @ctatedev.

Why it mattersFaster browser start-up means quicker, cheaper agent runs when they need to check a website. If your developer's agents browse the web, this is a free upgrade.

Sourcesx.com post ↗ · github.com ↗

Nace's Drex: small 'decision models' that only pick from your options

Nace.AI's Drex models take a case and a list of options and say how likely each one is, answering only from the options you give. Nace says Drex Max v1.1 now ranks first of 19 on the Artificial Analysis Decision Index, which is its own announcement; its docs list pricing at $0.12 per million input tokens. Shared by @dr_cintas.

Why it mattersFor yes/no or pick-a-category jobs, like sorting invoices or routing support tickets, a small constrained model can be cheaper and more predictable than a general chatbot. Test it on 50 of your own past cases first.

Sourcesx.com post ↗ · x.com post ↗ · nace.ai ↗

Matt Pocock is building /animatic, skills that turn work into short videos

@mattpocockuk previewed two agent skills he's been using: /setup-animatic, which picks and installs the video tooling, and /animatic, which makes an upload-ready video explaining a piece of work. His aim is for every important code change to come with a short video showing how it works. They aren't released yet.

Why it mattersA two-minute video often explains a change better than a page of notes, for clients and teammates alike. Worth watching for if you hand off work regularly.

Sourcesx.com post ↗

Greg Isenberg's 49-step playbook for an AI-run company

@gregisenberg posted a long checklist for running a company on agents, bookmarked by more than 1,600 people. The core: a 'second brain' folder of plain notes on how you think and decide; a company brain with customers, pricing and past decisions; a facts file of things agents may say about you, to stop made-up claims; recorded calls feeding it all; one skill per job; and a weekly review where an agent rewrites playbooks from the results and you approve the changes. He names many tools, including Obsidian, Wispr Flow, Granola, Orgo and Grok Bot; the list is his opinion, not a tested stack.

Why it mattersYou don't need all 49 steps. The facts file and the voice file are quick wins: two short documents that make every AI draft about your business more accurate and more like you.

Sourcesx.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.