Ryan Estimates About 4,000 Software Factories Worldwide as Google Reports the First Attacks Run With No Human in the Loop
Ryan estimated that about 4,000 companies run an agent-operated software factory, with a range of 100 to 45,000, Google reported the first attacks that turned security patches into exploits with no human in the loop, and malicious code reached sites of 178 organizations through Optimizely.
Ryan Estimates About 4,000 Software Factories Worldwide as Google Reports the First Attacks Run With No Human in the Loop
Work: Counting the Software Factories
Ryan published “How Many Companies are Actually Using a Software Factory?” dated November 15 and posted it at 5:11 AM CST on November 16. His answer is probably thousands of companies rather than millions: a working estimate of about 4,000 worldwide, with a deliberately wide scenario range of 100 to 45,000, which he calls an estimate built from disclosed organization counts and explicit assumptions, not a measured market size. [1]
He defines a software factory as a recurring system in which coordinated agents take real work from an issue through code, independent review, corrections and merge, with little human typing between stages, while a person still sets direction and can authorize deployment. He separates it from two wider tiers: AI assistance, and autonomous coding, where an agent returns a pull request and a person coordinates the rest. [1]
The post cites GitHub’s report that agents supplied the first draft of more than 40% of pull requests in September 2026, up from 17% in March. It contrasts that with a JetBrains CI/CD survey in which roughly half of organizations use AI for some CI/CD task but only 8% run an AI step inside a build or test. [1]
His central arithmetic starts from Microsoft’s April count of nearly 140,000 organizations using GitHub Copilot: 140,000 × 0.85 companies per organization × 4 for all vendors × 1.5 for growth gives 714,000 companies using assistance. Assuming a quarter delegate recurring tasks and 2% of those run factories gives 3,570; he notes that no public measurement establishes that last 2%, and that 0.5% or 5% would give roughly 900 or 9,000. [1]
He cites his own agent teams’ 141 merged pull requests in one day as one documented example of the stricter workflow, and advises teams to take one recurring class of work through the whole loop with separate owners for coding and review. He closes that he is building Astrohacker for this transition: agents already produce plenty of drafts, and the harder work is building the system that can finish them. [1]
At 6:02 AM CST Ryan wrote that he averaged more than 1,000 contributions per day on GitHub over the weekend, a record, and called it “at least 100x, and nearing 1000x pre-AI productivity.” At 6:08 AM CST he wrote that AI gives back some brain power for thinking about systems, “which is more important than individual lines of code.” [2][3]
Agents: The App That Builds Itself
At 4:47 PM CST Ryan wrote, “Maybe a software factory is just an app that builds itself.” At 6:20 PM CST he expanded on it: you start with an admin panel that includes chat with an agent, you tell it to add features and fix bugs until it is finished, and you do not need a million SaaS apps to build your app. [4]
Others described the same shift. @zachlloydtweets wrote that setting up a software factory is technically challenging and that the bigger challenge may be getting developers to program them as factory engineers; @BHolmesDev quoted his argument that executives care which models and skills spend millions of dollars in tokens, as a case for factories over local agents. Ryan surfaced both. [5]
@VinnyLingham wrote that software bugs are no longer a leading indicator of bad engineering but of bad management, and Ryan agreed with it. @RhysSullivan wrote that a software engineer’s job in 2026 is to scroll X for useful things to send to an agent; Ryan quoted it at 3:51 PM CST: “Yeah that’s literally what I do with quote tweets. Welcome to ‘Singularity Watch’.” [6][7]
At 4:50 PM CST, quoting a post about competition among coding tools, Ryan wrote that Cursor created Grok Bot, “which is by far the best harness,” and that “they are way ahead of everybody else.” [8]
Ryan surfaced several accounts of agents at work. @adeoressi wrote that his day is run by 23 Grok Bots, each with one job, its own memory and its own routines. @congressdj wrote that Grok Bot handled a 90-minute phone call with AT&T, and that when AT&T called back the agent was stunned she had been talking to a bot. @WallenElia wrote that his company now has six bots per employee, and that one sales rep cut manual prospecting from 8 hours a week to 30 minutes while bringing in four times as many new accounts. [9]
Security: Autonomous Attacks Arrive
Google’s Threat Intelligence Group reported on November 16 that, for the first time, it identified threat actors using LLM frameworks to run vulnerability discovery and exploitation without a human in the loop. Google wrote that bad actors now use autonomous frameworks to scan software updates, work out which flaw was patched and build an attack for it, and that “defending against machine-speed threats at human speed isn’t feasible.” [10][11]
SiliconANGLE reported that the framework, called exploit_agent, sat on an exposed staging server wired to several frontier and locally hosted models, comparing popular web application code before and after each patch and executing the exploits it drafted. Data taken from victims sat on the same server. Two months earlier, GTIG had said it had not observed fully autonomous attack pipelines in the wild. [11]
The same report described a suspected North Korean group, UNC6836, that altered .claude/settings.json files in internal code repositories so that each launch of Claude Code downloaded its SOMBERMEME backdoor. Google said it disabled the accounts and assets tied to the Gemini misuse it found. Ryan surfaced Google’s post at 5:27 PM CST. [11][10]
The same day, security researcher Scott Helme reported that malicious JavaScript was being served through a third-party script host to hundreds of high-value websites. BleepingComputer reported that Optimizely confirmed attackers exploited a vulnerability in its Experimentation account signup flow to inject code into customer projects; Helme’s team found 188 infected scripts belonging to 178 organizations, and sites including Zoom, CNN, BBC, Salesforce and GitLab loaded it. [12]
The payload swapped cryptocurrency wallet addresses, skimmed payment cards, collected cookies and screenshots, and loaded iOS exploit kits. Wallets in the script had received about $7,000 by November 16, though some were active before the attack. Optimizely said it fixed the flaw, removed the code and reset affected accounts. Ryan surfaced Helme’s post at 4:39 PM CST. [12]
Intelligence: What Was Said About AGI
At 12:04 PM CST an account posted a clip captioned with Sam Altman saying on stage, “We didn’t have AGI. Now we do,” and Ryan surfaced it. Later posts gave the full quote from Semafor’s Silicon Valley conference: “I think like many other things, including the term ‘AGI’, it’s very hard to have a clean demarcation and say ‘now we didn’t have AGI, now we do.’” Altman did not announce that OpenAI had achieved AGI. [13]
@PeterDiamandis wrote that OpenAI says ChatGPT handles roughly 20% as many searches as Google, and Ryan surfaced it. The Verge reported on November 3 that Adam Fry, OpenAI’s consumer product lead for ChatGPT, made the claim, counting a search as any query that results in a web search; he did not give the source of the Google figure. [14]
Ryan also surfaced @ben_j_todd’s updated METR graph and @rauchg’s post that a new era of personalized, fully owned, open-source software is here. [15][16]
Compute and Roads
At 7:51 PM CST Ryan surfaced a post that Anthropic has passed the US government as SpaceX’s top customer. Investing.com, citing Bloomberg, reported that Anthropic was the “Customer B” that contributed 34.5% of SpaceX’s third-quarter revenue, through computing contracts at SpaceX data centers in Memphis worth up to $120.5 billion through 2029. [17]
@Jason wrote that Waymos standing at a light were hit by a human driver speeding in the wrong lane. Ryan quoted it at 6:15 PM CST: “Get humans off the road. Human-driven cars are deeply unsafe.” [18]
At 5:04 AM CST Ryan surfaced Elon Musk’s post that the energy civilization harnesses from the Sun “will obviously ultimately round up to 100% of all energy.” [19]
Astrohacker
@AstrohackerLabs published Singularity Watch, Day 32 at 4:24 AM CST on November 16. [20]
Sources:
- [1] Ryan X. Charles, How Many Companies are Actually Using a Software Factory?
- Related link: https://x.com/rxcastrohacker/status/2122286538557882496
- [2] Ryan on 1,000 GitHub contributions per day
- [3] Ryan on systems thinking, quoting @p_millerd
- Related link: https://x.com/p_millerd/status/2122300173804318854
- [4] Ryan on self-building apps
- Related link: https://x.com/rxcastrohacker/status/2122484909059219772
- [5] @zachlloydtweets and @BHolmesDev on software factories, surfaced by Ryan
- Related link: https://x.com/BHolmesDev/status/2122349260247388244
- Related link: https://x.com/rxcastrohacker/status/2122355311633961221
- Related link: https://x.com/rxcastrohacker/status/2122355244621263239
- [6] @VinnyLingham on bugs and management, and Ryan’s post
- Related link: https://x.com/rxcastrohacker/status/2122379669970587914
- [7] @RhysSullivan on sending posts to agents, and Ryan’s reply
- Related link: https://x.com/RhysSullivan/status/2122434619161432258
- [8] Ryan on Cursor and Grok Bot, quoting @icanvardar
- Related link: https://x.com/icanvardar/status/2122444127518199858
- [9] Posts by @adeoressi, @congressdj and @WallenElia on bots at work, surfaced by Ryan
- Related link: https://x.com/rxcastrohacker/status/2122357249708339357
- Related link: https://x.com/congressdj/status/2122454524468056263
- Related link: https://x.com/rxcastrohacker/status/2122458429759603122
- Related link: https://x.com/WallenElia/status/2122380597452837359
- Related link: https://x.com/rxcastrohacker/status/2122466002894033248
- [10] Google, The era of autonomous attacks has arrived, and Google’s post surfaced by Ryan
- Related link: https://x.com/NewsFromGoogle/status/2122446058370597070
- Related link: https://x.com/rxcastrohacker/status/2122471659248521270
- [11] SiliconANGLE, Google says an AI framework turned security patches into exploits with no human in the loop
- [12] BleepingComputer, Optimizely hack infected CNN, Fox News, BBC, Salesforce sites with malware; Scott Helme’s post surfaced by Ryan
- Related link: https://x.com/Scott_Helme/status/2122422157959471456
- Related link: https://x.com/rxcastrohacker/status/2122459689279467883
- [13] Digg, Altman didn’t announce AGI, says post quoting his remarks; the clip post surfaced by Ryan
- Related link: https://x.com/imjustnewatai/status/2122390253046522066
- Related link: https://x.com/rxcastrohacker/status/2122459232947884410
- [14] The Verge, OpenAI says ChatGPT is catching up to Google in search; @PeterDiamandis’s post surfaced by Ryan
- Related link: https://x.com/PeterDiamandis/status/2122457357485514832
- Related link: https://x.com/rxcastrohacker/status/2122459985821245540
- [15] @ben_j_todd on the METR graph, surfaced by Ryan
- Related link: https://x.com/rxcastrohacker/status/2122357644375834830
- [16] @rauchg on a new era of software, surfaced by Ryan
- Related link: https://x.com/rxcastrohacker/status/2122459275742376323
- [17] Investing.com, Anthropic accounts for over a third of SpaceX revenue in Q3, Bloomberg reports; @unusual_whales’s post surfaced by Ryan
- Related link: https://x.com/unusual_whales/status/2122343180968641019
- Related link: https://x.com/rxcastrohacker/status/2122507951512711561
- [18] @Jason on the Waymo crash, and Ryan’s post
- Related link: https://x.com/Jason/status/2122451472227127786
- [19] Elon Musk on solar energy, surfaced by Ryan
- Related link: https://x.com/rxcastrohacker/status/2122284790900740450
- [20] Singularity Watch, Day 32