Parallel cursors, an $11.5M seed round, and a Claude Code flag with no off switch

X Post Facto is our weekly pause on the feed: a look back at the AI posts that stood out, what they said, and why they’re worth a closer read.

Across 25–31 October 2026, this set runs from agents on the computer to questions about what those agents are allowed to do. One post pointed coding agents at the whole computer, in parallel. A code editor was renamed around running many of them, with an $11.5M seed round attached. A model at ~30B parameters was announced with a teacher-first plan for small agents. Anthropic described a check that flags potential misuse right after Claude replies. A separate post said Claude Code had picked up a rule sent from Anthropic’s servers, with no setting to turn it off. Ledger asked who is in control when agents prepare trades, payments, and purchases. The smaller notes are a GTA mission, living wallpapers, Grok dashboards, and a plant.

Parallel cursors, while you keep working

Chris Tate introduced agent-desktop as “agent-browser, for your whole computer.” The offer is “Computer Use for any coding agent,” on macOS, Windows, and Linux. It “Runs in bg while you keep working.” Agents use the computer in parallel, “each with its own cursor.”

Background use is the practical pitch. Parallel cursors, on a machine you are still using, stay a convenience only if you can tell which one moved what.

Armies of agents, and seed money

Paradev, formerly Superset, put the new name in one line: “The code editor that helps engineers run armies of coding agents.” It said it had just raised an $11.5M seed round led by @usv with @ycombinator “to redefine how software is written.”

“Armies” is a poster word. The sharper shift is the editor’s job, as they frame it: not only where you type, but where you run agents. A seed round funds the attempt. It does not, by itself, redefine how software is written.

Build the teacher first

Nanbeige announced Nanbeige4.3-33B and led with the method. To push the performance ceiling of small agents, “our strategy in Nanbeige4.3 starts with building a powerful teacher first.” The model, they said, delivers “SOTA agentic performance among open models at ~30B parameters,” and the post points to strong agentic results.

The sequence is the usable idea. In their telling, a higher ceiling for small agents starts with a powerful teacher. The SOTA line is the claim they hang on that order, among open models at that size.

A flag with no off switch

chasm wrote that Anthropic is cracking down on video game rewrites. Claude Code, he said, picked up a new rule around Oct 27, sent from Anthropic’s servers under a flag called “tengu_heron_brook”. “There's no setting to turn it off.”

He calls what follows “The full rule,” and the words begin “Do build original games, games in a genre, and games,” with the line running on from there. Beside a headline about rewrites, that opening is about what to build. The part with consequences, in his account, is the delivery: from Anthropic’s servers, with no setting to turn it off. That is not a preference you flip before lunch.

A preliminary grade, after the reply

Anthropic said it has trained Claude to flag potential misuse in conversations, “a technique we call self-assessment.” “Right after it responds, Claude gives a preliminary safety assessment of the conversation.” Since it began using self-assessment in June, Anthropic said, it has surfaced misuse.

The timing is the design choice worth keeping. The check comes after the response, and they call it preliminary. It is a different announcement from the flag above: Anthropic describing a safety pass on the conversation, not the same post as the rewrite rule.

Prepare it, then ask who is in control

Ledger led with the exciting half. “AI agents can now prepare trades, payments, purchases, and more for us.” Agents acting on our behalf, the post says, raise the question of “who's in control?”

It names @iancr, Chief Human Agency Officer, @pgauthier, Chairman and CEO, and @P3b7, CTO. Chief Human Agency Officer, beside that question, is hard to read as filler. “Prepare” is doing careful work too: get the trade, the payment, or the purchase ready, and the control question is still open.

Vibecode the sequel from inside the last game

Aravind Srinivas asked Perplexity Computer to design a GTA V mission “where you vibecode GTA VI from an in-game phone while the devs call the cops on you for shipping GTA VI before them.” He said it went through the FiveM docs — natives, ped models, server convars.

The premise is the demo: ship the sequel from a phone inside the last game, and have the developers call the cops for getting there first. What keeps it from being only a joke is the documentation he says the agent opened. Natives, ped models, server convars — a modder’s reading list, not a mood board.

Eight wallpapers, labeled a fan concept

Marko Vineli posted a /theme for Claude Code “but every wallpaper is alive”: eight of them, “pick yours,” labeled a “fan concept.”

Eight to choose from, and the wallpaper is meant to be alive. A fan concept with a pick-one invitation is a small, cheerful note beside the week’s flags and control questions.

Pages, dashboards, and a usage claim

Alex Finn said Grok Bot just had its biggest update yet. It introduces pages and dashboards, “2 things that solve Grok's BIGGEST problem.” He says the video covers five dashboards that make Grok Bot “SO much nicer to use,” including one that has saved him “a TON of usage.”

He is specific about the new surfaces and about a saving from his own use. “Biggest update” and “BIGGEST problem” do a lot of lifting for pages and dashboards.

A component list, then the plant

santi had wanted, for a while, to learn about hardware, so he started talking with Claude about what project he could do. They landed on one he liked: “medir la humedad de la tierra de mi planta para saber cuándo regarla.” Measure the moisture in his plant’s soil, to know when to water it. The first thing Claude did was put together a list of the components.

Next to talk of agent armies, one plant is a kinder brief. The useful move is the order: a project he liked, then a component list. For hardware learned in a chat, that is a standard worth stealing — a first reply the project can be held to.

Start with the post that matches the question in front of you. Tate’s agent-desktop introduction and Paradev’s rename are the ones about agents on the machine. Anthropic on self-assessment, chasm on the flag, and Ledger on control are three different limit stories; do not collapse them into one policy. For a smaller start, Srinivas’s GTA brief and santi’s plant are the ones you can actually try.