Good day . People are deleting the personal AI agents they just installed, Anthropic is handing startups a free year of Claude, and Google cut its image-model prices in half. We also get a plain-English tour of Codemode, and Google Cloud’s pitch to haul your mainframe into the sky. Let’s get into it.
Early Users Are Deleting Their AI Agents Business Insider
What happened: Business Insider talked to four early adopters of Meta’s Muse and a startup agent called Instinct who deleted or restricted the tools, with dozens more raising the same worries on social media. The triggers were privacy scares and blunders: agents reading one-time login codes out of Gmail without asking, making up personal details from a document nobody sent, and setting off a carrier-account login prompt that appeared to come from Iran.
Why it matters: A personal agent only works if you hand it your inbox, calendar and payment methods. Scott Persinger, CTO of travel platform BizTrip, put the stakes plainly: “Access to email is everything — via password resets you could probably access my whole life.” Sunday’s brief covered Muse profiling your friends, and this is the user-side reaction.
What everyone’s saying: Growth is still pointed up. The Information reports Muse has more than 3 million weekly users, and Instinct’s founder says his invite-only service is growing roughly 10% a day. Meta says Muse has “first-of-its-kind privacy, safety, and security protections” and asks before big actions like purchases. Instinct didn’t respond to Business Insider. YieldClub’s CEO called it “the wild west.”
My read between the lines: Look at who’s leaving. These are founders and CTOs, the people who most want agents to work, and they’re still trimming the permissions. The ones who keep using the tools are withholding passwords and skipping the inbox. Right now the product is “agent with the access turned down,” which is a long way from the pitch.
📖 Further reading: A Stranger Sent Me an Invite to an AI Nobody Has Written About — my deep-dive on Instinct, the invite-only agent at the center of this story.
Everyone wants an agent in their inbox, and almost nobody can tell you what it did with the keys. Viktor lives in your team’s Slack instead, connected to 3,000+ of your tools, and does the work you can describe: the report, the dashboard, the code, the campaign. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor →
Anthropic Hands Startups a Free Year CNBC
What happened: On Tuesday Anthropic expanded its Claude Startups program, which launched in May and, per Anthropic, already has thousands of companies in it. TechCrunch reports members get a free year of Claude Team with up to five premium seats, a one-time $1,000 in API credits, access to the Claude Marketplace and virtual office hours with Anthropic’s applied AI team. Companies founded in the last five years or funded in the last two can apply.
Why it matters: Startups pick their tools early, and whatever they build on at the start tends to be what the product runs on later. Free seats and credits are how a vendor gets into a company before that company has a budget. If you’re a founder, this is real money off for the first year.
What everyone’s saying: CNBC frames it as Anthropic deepening ties with founders while it fends off OpenAI and Google, which both run their own startup programs, and as it gears up for what CNBC calls a widely expected blockbuster IPO. CNBC also notes the vast majority of Anthropic’s revenue comes from business customers.
My read between the lines: The headline number is “up to $45,000” in discounts and credits through the Claude Startup Stack, which Anthropic describes as benefits for the tools a startup runs on. The part that is clearly Anthropic’s own is the $1,000 and the seats. Both sides want the first year to feel generous and the second year to feel like habit.
📖 Further reading: Claude for Small Business Just Hired You 44 Employees. You Only Need Four of Them. — my paid look at what Anthropic is building for small companies, which is the crowd this program courts.
The daily Brief is free and always will be. The paywalled deep-dives are where I stay with one story long enough to be useful, like my dive on Instinct, the invite-only agent in today’s lead. Members get those plus the full archive. Become a member →
Nano Banana 2.1 Costs Half as Much The Decoder
What happened: Google released Nano Banana 2.1, a new image generation and editing model built on Gemini 3.6 Flash, on Tuesday. A 1K image drops from 6.70 cents to 3.36 cents through the API, and a 4K image from 15.10 cents to 7.56. It’s rolling out in the Gemini app, AI Mode in Search, Google AI Studio, Flow, Stitch, Google Ads and the Gemini Enterprise Platform.
Why it matters: If you build anything that makes pictures, every image is a line on a bill, so a price cut this size shows up in your margins. There’s a deadline too: Google shuts down the older
gemini-3.1-flash-imagemodel on October 29, so developers have a few weeks to move.What everyone’s saying: Google’s own preference tests give 2.1 in thinking mode a score of 1050, against 990 for Nano Banana 2 and 935 for Nano Banana Pro. The Decoder says it can top Pro by a wide margin on benchmarks, but in practice Pro often still makes noticeably better images. The reviewer’s horse-rides-astronaut test came back with the horse looking more like a pony.
My read between the lines: Google calls 2.1 “the more efficient counterpart” to Pro, which is the polite way to say it’s the cheaper one. Notice the sentence structure: it wins on Google’s tests and loses on a horse. For a thumbnail factory at half price, fine. For the picture you’re proud of, you’re probably still paying for Pro.
📖 Further reading: What Is Nano Banana—And Should You Be Excited or Skeptical? — my starter take on the model family, and a half-price tag is a good moment to ask that question again.
Codemode Lets AI Agents Write Their Own Tool Calls Armin Ronacher
What happened: Developer Armin Ronacher published a long explainer on Codemode, an idea in which an AI agent writes a short JavaScript program to call its tools, instead of calling them one at a time. The program runs in a locked box on the harness side, the software wrapped around the model. In his Pi coding agent that box is a small sandbox with no network, no file system and no timers. He says Cloudflare coined the name.
Why it matters: Plain command-line tools can only chain programs together, and some things an agent needs aren’t programs: reading an image, spinning up sub-agents, or calling an image or classifier model. Code mode lets the agent loop over 100 items and keep the bulky results out of the model’s context. It’s also how Pi 1.0 supports MCP, the standard plug for connecting tools to AI.
What everyone’s saying: Hacker News commenters were mostly skeptical. One called it “a very complex apparatus for little gain” and said tool calls could just be disguised as bash commands. Another stripped Codemode out of a prototype because when one embedded tool call failed, the whole script failed and the small model kept rewriting big chunks of code. A third asked how it differs from Claude writing little scripts, which it already does.
My read between the lines: Ronacher is unusually blunt about the limits. Codemode works “not amazingly well” with MCP today, because servers don’t return clean structured data and can’t handle large binary files. He also lists durability, images and smaller models as open problems. The last third of the post is a repair list, which is the part I’d trust.
Google Pitches AI Agents for Legacy Migration TokenPost
What happened: Google Cloud introduced Modernize, a bundle of migration services, AI-assisted assessment tools and infrastructure aimed at enterprises. It includes a Modernization Hub that analyzes code and maps dependencies in Java, .NET and mainframe applications, a generally available Agentic Quick Estimator that projects Compute Engine costs from VMware exports, and a public-preview agent for moving workloads from Amazon EKS to Google Kubernetes Engine.
Why it matters: Moving old systems like mainframes and VMware setups is the slow, expensive part of enterprise IT, and it’s the project nobody wants. If agents can size and map that work before humans start, the first quote gets cheaper and faster.
What everyone’s saying: Google points to customers: NetEase Games cut server costs by 40%, and Deutsche Börse Group cut aggregation latency by more than 50%. Those are Google’s numbers, from a launch announcement. The hardware side adds X5 virtual machines with up to 43 tebibytes of memory in one node.
My read between the lines: Read the savings claim closely. The 20%-plus licensing cut Google cites for Oracle and similar databases comes from the memory-heavy M4N virtual machines, not the AI. The agents estimate and map, the hardware does the saving, and the EKS-to-GKE agent is a way to pry workloads off a rival cloud. It’s still in preview.
That’s your AI Brief for Wednesday.
—Nicholas from Artificially Intimidating
