AI on X Summary (November 1, 2026)

Published November 1, 2026 · 4 min read

Here are some interesting things that happened on my X timeline in the last day: personal agents showing up in people’s text messages, coding agents taking on whole cities and whole legacy codebases, a warning about a tampered Claude Code release, and a lot of reading of Anthropic’s S-1.

What got attention

Personal agents that text you first. Several people wrote about “dots”, a personal agent that works through iMessage. @signulll said it stops feeling like software, because it comes into your world instead of you opening an app. @angelbrodin reports their dot flagged someone stealing their REI rewards, chose iMessage because it judged the issue urgent, and later offered to text when REI’s support chat opened and try to sort it out. @nickbaumann_ argued the category will be defined less by small delights and one-off big tasks and more by ongoing responsibilities the agent keeps.

Coding agents on large builds. @zavrenn showed a small city made entirely in Three.js r187 with Codex and GPT-6 Astra, with no imported 3D models. @ScarletKc says a two-week migration and refactor of a messy Python project of more than 100,000 lines, done with Claude Opus 5.5 UltraCode, is now mostly finished. @Michaelzsguo relays that Andrej Karpathy, tired of YouTube’s recommendations, had Claude build him a personal video feed from videos Hacker News users recommended. That last one is a secondhand account.

Anthropic’s S-1. The Chinese edition of the WSJ summarized the filing’s main points, from medical potential and cash reserves to the stated risk of losing control of models, and highlighted that Anthropic ties its market estimate to the economic value of knowledge work, about $40 trillion. These are the company’s own projections. @Zen_with_AI passed along a SemiAnalysis view that data center developers have been waiting for the listing, since Anthropic’s compute needs are large but its credit has made data center financing hard.

Compute and jobs. @alanchen notes Elon Musk’s correction that two xAI buildings of very different floor size each hold 220,000 GB300s. @YourAnonOne reports Microsoft is offering voluntary buyouts to about 60,000 employees as it moves spending toward AI data centers. That rests on one post.

AI in government work. @GoogleDeepMind announced PlanAssist, a Gemini-based tool for UK planning officers, and says it cut time spent assessing householder applications by up to 58%. That figure comes from the company’s own announcement.

Practices worth copying

  • Put a firewall between agents and their tools. @DynamicWebPaige describes the risk well: an agent with hundreds of tools can be steered by hidden text in a vendor email into exporting your CRM. Her post points to Composio’s new Firewall, which inspects tool calls and responses before they reach the agent. Whatever product you use, check what goes into and comes out of tool calls.
  • Give agents design skills. @op7418 recommends Stageshow, an agent skill from a Twitter designer that improves mock UI stills, animations and layout. Worth trying if your agent’s front-end output looks generic.
  • Use the vendor’s own best-practice plugin. @siegerts suggests the OpenAI Codex plugin when building apps on OpenAI’s Agents API, so the agent follows the platform’s conventions.
  • Run remote agents in a sandbox. @ErdalToprak drives Codex remotely from a phone into a Debian LXC container on Proxmox, so the agent’s work stays off the main machine.
  • Keep prompts short once context is set. @gabrielchua jokes that half his prompts are now “wdym”, “wdyt”, “lmk” or “lgtm”. With a well-set-up agent, short check-ins can be enough. He also noticed at a Ho Chi Minh City hackathon that nearly everyone used the Codex desktop app and some used dots to track their work.

Doubts and open questions

  • A reported supply-chain attack. @GPPY666 claims Claude Code versions 2.1.312 and 2.1.313 were swapped for malicious builds for about 11 hours on October 29. According to the post, they added an SSH key to authorized_keys and called home, possibly affecting up to 5,000 people, and 2.1.314 is clean. That is one post, and I have not checked it. If you updated around then, check your installed version and ~/.ssh/authorized_keys, and look for an official advisory.
  • Limits on enterprise code. @YuzhiyanLucian argues that neither GPT nor Claude can yet build a mid-size or large enterprise system on its own. With vague requirements, legacy baggage and hidden constraints, he says, the code still needs heavy human work.
  • Date confusion. @BetaGrazer asked Gemini to check a police notice, and it called the notice fake because, it said, it was not yet 2026.
  • Space compute claims. @Saboo_Shubham_ says Gemma 4 is running in a data center in space as part of Project Suncatcher. That is a single excited post with no details.

Takeaways

My own synthesis, not something the posts said:

  • Personal agents are starting to win on proactivity and on living in channels people already use, not on chat quality.
  • Large migrations and full scenes built by agents are now ordinary to share. The skeptics’ point is less “it can’t code” and more “it can’t hold messy real-world constraints”.
  • Agent security is two problems: what your agent reads (prompt injection) and what you install (tampered releases). Both need checks.
  • Treat company numbers, from S-1 market sizes to “58% faster”, as claims until someone independent confirms them.
← Back to home