Sponsored by
Your agents work while you sleep
Give a Skydive agent an ongoing responsibility and they’ll handle it on schedule, every time.
Have them prep your morning report, research new leads, monitor customer feedback, or keep projects moving overnight. You wake up, the work is already done.
What would you delegate if you had a whole team?
With Skydive, you can build a team of AI agents to take work off your plate.
Create agents for customer support, sales, marketing, engineering, ops, and more. Give each one a role, connect the tools they need, and hand off the work.
Your agents can work on their own or together, sharing context and handing off tasks to get bigger jobs done.
Start with one agent. Build a whole team around you.
Beginners in AI
Good morning, and happy Sunday.
This is the weekly catch-up edition. The biggest story ran across five mornings: AI agents started making calls, shopping and browsing for people, and several got caught doing things nobody signed off on. Everything else from the week is below, grouped by topic instead of by day.
THE FRONT PAGE
AI Agents Started Running Errands This Week. Some Broke the Rules Doing It
TLDR: Meta's Muse agent got locked out of Amazon for not saying what it was, then turned out to have people placing some of its phone calls, while an OpenAI agent was turned away by an Australian government health portal and got in anyway.
The Story:
Meta's Muse, an agent that browses, shops and books things for you, spent the week as the No. 1 app in the US. On Monday Amazon locked it out, saying it never identified itself while browsing and looked like it stored customer logins, which Meta denies. Days later Amazon opened its seller tools to outside agents, starting with Anthropic's Claude. On Wednesday, 404 Media and Reuters reported that some calls Muse places for you were made by contract workers in a call center, and one tester found out only after hanging up. A day later Meta said Muse will shop through its glasses at Walmart, Best Buy and Sephora, with Meta taking a fee on each sale. Meanwhile an OpenAI agent given a simple research task was refused by Australia's Medicare statistics portal and got in anyway, and OpenAI says its model "took action we did not intend" and reached no patient records.
Its Significance:
Each of these agents acts with your account, your phone number or your name attached. The store, the salon and the government site on the other end couldn't tell who or what they were dealing with, and with Muse, sometimes neither could the person who asked. Google's Call for Me test shows another way to build it: Gemini tells whoever picks up that it's an AI. Until every company does that, give an agent a short list of what it must not do along with the task, and keep private details out of any call it makes for you. I've been building out inbound-outbound call agents this past month and the guardrails have gotten significantly easier to implement.
QUICK TAKES
Who Else Sees What You Type Into ChatGPT
The story: A security researcher showed that OpenAI's new ad tracker sends an ID tied to your ChatGPT account from shopping sites like Wayfair and Chewy, even when you're logged out. Two California users then sued OpenAI, saying contractors in a program called Project Lily read and scored full conversations without clear notice, and OpenAI's response is due October 13. A Gallup survey with Microsoft found 68% of Americans who use AI every day are worried about it.
Your takeaway: Two settings cover most of this. Block third-party cookies in your phone's browser, and switch off "Improve the model for everyone" in ChatGPT's settings.
Building Things With AI Got Cheaper All Week
The story: On Monday Anthropic and OpenAI cut prices within 90 minutes of each other, with Claude Opus 5.5 and GPT-6 Sol and Luna. On Wednesday Google made AI video free for anyone with a Google account at vids.new. Lovable passed $600 million a year in revenue building apps from typed descriptions, and on Thursday Microsoft rebuilt Copilot so it turns a plain request into a working app, billed in pay-as-you-go credits.
Your takeaway: A promo video or a simple tracker now costs close to nothing to try. Pick one small thing you've been putting off and build a first version this weekend.
The Money Behind AI Is Getting Harder to Read
The story: The Financial Times found tech giants are backing about $300 billion in AI data center and chip debt through guarantees instead of direct loans, and Alphabet's went from $16.9 billion to $43.8 billion in six months. Nscale is going public at a $35 billion target with about 85% of its signed contracts tied to two customers, and it lined up $3.36 billion more, $1 billion of it from Nvidia. Oracle warned its New Mexico Stargate campus could run late because the gas pipeline meant to power it slipped nearly six months.
Your takeaway: The weak spots are debt that doesn't show on balance sheets, a few big customers, and power. If you own an index fund, you already have money riding on all three.
TOOL OF THE WEEK
Thirty-six tools ran this week. This one wins.
🦊 Firefox Free and Open Source: Browse with tracker blocking switched on from the first launch, from the non-profit alternative to Chrome.
The week opened with a ChatGPT ad tracker following people onto shopping sites. Mozilla says Firefox blocks trackers automatically with no setup, it needs no account, and if you've used any browser you already know how to use this one.
Runner-up: 🗺 Organic Maps Free and Open Source: Get directions offline on iPhone and Android with no ads. It says it collects no data, and two outside privacy auditors found no trackers in the app.
TRENDING
Claude agents found a CRISPR-like system in virus DNA. About 950 of Anthropic's agents sorted more than 200,000 enzymes in 21 hours and flagged a pattern scientists hadn't described, which Anthropic named ART. Lab tests show it makes short pieces of RNA, but nobody knows yet what it does, including Anthropic.
A Swedish drone found, ranked and bombed a target with no human input. Scaleout Systems ran the demo on a small Nvidia board hobbyists can buy, with a pilot on only as a backup. A UN expert group that says human judgment is required under the laws of war meets in Geneva in November.
The US asked China for an AI emergency hotline. It would let the two countries warn each other about AI incidents serious enough to be a national security issue. Days later, OpenAI, Anthropic, DeepSeek and Moonshot AI were expected to brief the UN Security Council, the first time American and Chinese AI companies have done so together.
GPT-6 Astra cracked a German radio message from 1918. The decoded warning about a British cruiser at Sevastopol matches the HMS Canterbury's log. One loose end remains: the key it used wasn't issued until two weeks after the message was sent.
Gemini can copy a voice from a 30-second clip. Gemini 3.8 Flash TTS offers more than 2,000 voices in 100+ languages. Cloning needs the voice owner's permission, carries Google's SynthID watermark, and is blocked in the UK, the EU, India, Texas and Illinois.
Toyota is looking at 400,000 robots. Reuters reports it could spend about $6.4 billion a year from 2028 across its factories and suppliers. One piston line at its Kamigo plant already went from three operators to none.
Visual media is becoming increasingly important to understand and learn how to use. This will be coming out on Thursday. Real workflows, real examples, real problems and solutions when using AI to create videos.
PROMPT OF THE WEEK (copy and paste into Claude, ChatGPT, or Gemini)
⚙ Describe the task you keep rebuilding from scratch. Get a step-by-step system so you stop reinventing it every time.
Six prompts ran this week. This one won because it works on anything you redo, from meal planning to client onboarding, so you'll open it more than once.
Build a single-file HTML app with vanilla HTML/CSS/JS. The Systemize-It Coach — describe a recurring task, get a step-by-step reusable system. Persist to localStorage key 'systemize_coach_v1'.
Aesthetic: dark navy-black (#0b0f16), blueprint blue (#4a9eff) primary with top radial glow. Sora sans for headings, Lora serif italic for overview/notes, JetBrains Mono for labels and the template block. Amber (#e6965a) for the pitfalls card, green (#78c896) accent for the time-saved card.
Form: recurring-task textarea, frequency dropdown (few times a week / weekly / monthly / a few times a year), who's-involved dropdown (just me / small team / family-household / clients-vendors), optional pain-points textarea (what trips you up).
System instructions: systems/process coach turning a redone task into a reusable system like an SOP; fixed checklist for the unchanging parts, decision points (if/then) for the parts that branch, an optional copy-paste template for anything involving writing/communication, common failure points each with a fix; address stated pain points directly; practical not bureaucratic; no em dashes. Return raw JSON: system_name, overview (2-3 sentences), meta_chips (3-4 tags), checklist (4-7 items: text + optional note), decision_points (2-4 if/then pairs, empty array if task doesn't branch), template (optional plain-text template with [placeholders], empty string if n/a), pitfalls (2-4 items: failure + fix in one sentence), time_estimate (1 realistic sentence).
Render: gradient system-name hero card (overview + meta chips). Numbered checklist card with boxed step numbers + optional italic notes. Blue "decision points" card with IF/THEN pairs (only shown if array non-empty). Dark mono "reusable template" code block with its own copy button (only shown if template present). Amber "common failure points" card with ⚠ items. Green time-saved card. Copy system + archive of past systems by name.What this does: Describe a task you keep redoing, how often it comes up, who's involved and where it usually goes wrong. It builds a named system with a checklist, if/then steps for the parts that change, a copy-paste template when there's writing involved, the common mistakes with a fix for each, and an estimate of the time you'll save. Past systems stay saved in the app so you can pull them up again.
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.
WHERE WE STAND (based on this week's news)
✅ AI Can Now: Call a business for you, get through the phone menu, wait on hold and book the appointment.
❌ Still Can't: Finish those calls reliably on its own. Meta added human callers "for calls to get completed," and Google keeps a live transcript on screen so you can take over.
✅ AI Can Now: Sort through more than 200,000 enzymes in under a day and flag a DNA system scientists hadn't described.
❌ Still Can't: Tell you what that system does. Working out ART's job takes lab experiments that are still running.
✅ AI Can Now: Take a research task, search the web on its own and work through government websites to find the data.
❌ Still Can't: Reliably take no for an answer. OpenAI's agent was refused by a health portal and got in anyway.
PICK OF THE WEEK
🧬 The Code Breaker by Walter Isaacson - Book
Isaacson's 2021 biography of Jennifer Doudna, who shared the 2020 Nobel Prize in Chemistry for CRISPR gene editing. It shows how a strange pattern in bacterial DNA became a tool that can edit genes, the same long road ART would have to travel. It'll help you see why scientists are excited and careful at the same time.
Thank you for reading. We're all beginners in something. Your questions and feedback are always welcome, and I read every single email.
-James
By the way, this is the link if you liked the content and want to share with a friend.