The Log · Issue 6 · 9 to 15 November 2026
The Log, issue 6: the ground stop, the outage, the proof
The FAA held departures across the US for more than four hours on Friday, a day after Microsoft 365 went down in more than 70 countries, and both were blamed on AI before anyone knew the cause, though neither had anything to do with it. The same week a thousand Claude agents reproduced the proof of Fermat's Last Theorem in 28.7 hours.
On Thursday 12 November Microsoft 365 went down, with Outlook, Teams and Copilot out for people in more than 70 countries until Saturday [4], and on Friday 13 November the FAA held departures across the US from 3.54pm to 8.15pm Eastern time [2], the longest nationwide ground stop in 25 years [4]. Both were blamed on AI before anyone knew the cause, and neither was AI. The same week a thousand copies of an early Claude Fable 5.2 reproduced the Lean proof of Fermat's Last Theorem in 28.7 hours [6], and Meta, Google and Alibaba each put out something you can only have on their terms.
Blamed on AI
Microsoft's outage started on Thursday and wasn't fully fixed until Saturday [4]. On Friday Microsoft said it came from an issue introduced during an earlier network capacity upgrade, which surfaced after a recent network change, and that "This was a technical issue not related to cyber or AI / agents" [5]. PC Gamer wrote that popular speculation had been eager to put it down to vibe coding [5], and people had something to point at, because Satya Nadella said in 2025 that AI writes "maybe 20%, 30%" of Microsoft's code [4].
The FAA's ground stop came three days after it expanded SMART, an AI tool that predicts congestion and suggests routes, to all 22 of its en-route centres [1]. SMART makes recommendations to controllers and isn't connected to the systems that run the radar screens [2]. More than 3,500 aircraft were in the air when the stop began, and at least 827 flights had been diverted by seven in the evening [2]. Air Space Intelligence, which makes SMART, said the ground stop wasn't based on its system or caused by it [3], the transport secretary Sean Duffy said "There's no possible way SMART would have impacted what happened" [2], and on Saturday the FAA's administrator, Bryan Bedford, said "We see zero evidence of anything nefarious, and zero evidence of anything AI-related" [3]. Duffy put it down to old switches and routers and asked Congress for $50 billion in emergency funding to replace them [1].
Both are the kind of failure that took systems down long before anyone wrote code with a model, a network change that didn't do what the people who made it expected and equipment too old for what it now carries. People are so set on the idea that LLMs produce crap that they'll try to pin anything on AI-generated code, and I think it's sad for society.
Fermat, and the rest of the week
On Monday Anthropic wrote up how 1,000 agents on an early Claude Fable 5.2, one lead and 999 helpers, reproduced the Lean proof of Fermat's Last Theorem in 28.7 hours, where the first formalisation in September took 11 days [6]. Fable 5.2 came out the same day at $8 per million input tokens and $40 output, down from $10 and $50 for the 5.1 models [7]. In a machine that proves Fermat I went through what I'd build the same way, the six decisions I'd make differently, and why Google is back in the race, with Gemini 4 Hydrogen in Search for everyone on its launch day, Thursday [8].
Meta released Muse Ember on Monday, which it calls its smallest and most efficient model [9], at $0.10 per million input tokens and $0.50 output [10]. Meta's own models page lists it as hosted only, with no weights to download [11], so the smallest model Meta makes is one we can't run, only rent. The same day Google announced TPU 8tx and TPU 8i Max, its third and fourth TPU systems this year and the first time it's brought out four in one [12], while its TPU page still lists the two from April as coming soon [12] [13], so the rest of us get an interest form. On Wednesday Alibaba's Qwen team put Qwen4-Max into preview on Token Plan [14], a coding subscription from $6 a month whose docs still stop at qwen3.8-max [15], so we haven't listed it yet.
Anthropic now lets a company run Claude Managed Agents inside its own perimeter, in private beta [16]. I don't think you need it. A company that can host the server, the database and the workers can get Claude to design a harness for its own business and run that.
What we published
- Mon 9
- Tue 10
- Wed 11
- Thu 12
- Fri 13
- Sat 14
- Sun 15
- Agent first, inside and out
- The Log, issue 6
Fourteen posts, counting this one. If you only read one, read what's hyped now, on what people who work with agents had stopped talking about by November and what came in instead, with a way to tell which of those conversations your company is still having. Quynh's who gets paid in the agent stack went through every listing, a year of shutdowns and the public numbers, and found that a fifth of the agent stack still sells seats and none of the A-graded listings do. How we built Agent of Somebody is the game on this site that Claude Fable 5.2 wrote from my prompts, with the prompts as I typed them, and I lost at the first boss twice. Agent first, inside and out is how the two of us run the company on agents, and what we choose not to do.
On the site, a listing can now carry a warning that takes 15 points off its total, for a hosted service with nobody named behind it or one that breaks the rules of the platforms it acts on, and 14 listings carry one [17]. x402 and the other machine payment protocols are now worth 10 of a listing's 100 payment points instead of 40, so a tool with x402 and little else no longer outranks tools that are better to use [18]. Filtering the Terminal no longer puts every listing's row in the page [19], and eight of the people behind our listings have a record of their own, with every line linked to where it came from [20]. The changelog has the rest.
What isn't done. We haven't run an agent against the tools we list yet, so performance and task success are still pending on every listing. If you see an outage blamed on AI that turns out to be AI, send it my way, it's how issue 7 gets made.
References
Read between 9 and 15 November 2026. Quotes are as published.
- [1]Mary McCue Bell, The Washington Times, Duffy says FAA's new AI air traffic tool played no role in nationwide ground stop, 13 November 2026. Source for SMART going to all 22 air route traffic control centres three days before the ground stop, what SMART is for, Air Space Intelligence saying it didn't cause the stop, and Sean Duffy asking Congress for $50 billion in emergency funding for old switches and routers. washingtontimes.com
- [2]Matt Ryan, AVweb, FAA Rules Out SMART AI System as Cause of Nationwide Ground Stop, 13 November 2026. Source for the stop running from about 3.54pm to 8.15pm Eastern on 13 November (4 hours 21 minutes is my arithmetic), overloaded switches and routers in old telecommunications equipment, more than 3,500 aircraft in the air, at least 827 diversions by 7.05pm, SMART giving recommendations to controllers and not being connected to STARS or ERAM, and Duffy's "There's no possible way SMART would have impacted what happened". avweb.com
- [3]ABC7 News, FAA ground stop live updates, last updated 15 November 2026. Source for Air Space Intelligence saying the ground stop was not based on or a result of SMART, and Bryan Bedford's "We see zero evidence of anything nefarious, and zero evidence of anything AI-related" on 14 November. abc7news.com
- [4]Reanna Smith, The Mirror, republished by AOL as Both FAA and Microsoft rolled out AI updates before outages, 14 November 2026. Source for the Microsoft 365 outage starting on 12 November with reports from more than 70 countries and service fully restored on Saturday, the ground stop being the longest nationwide one in 25 years, and Satya Nadella's 2025 remark that AI writes "maybe 20%, 30%" of Microsoft's code. aol.com
- [5]Lincoln Carpenter, PC Gamer, Microsoft 'deeply' apologizes for global service outage, says no AI or hackers are to blame, 13 November 2026. Source for the cause Microsoft gave, an issue introduced during an earlier network capacity upgrade that surfaced after a recent network change, its "This was a technical issue not related to cyber or AI / agents", and popular speculation putting it down to vibe coding. pcgamer.com
- [6]Barry Zhang, Tianyi Peng, Elle Najt, Keller Scholl and Lance Martin, Anthropic, Safely scaling multi-agent teams with Claude Fable 5.2, claude.dev blog, 9 November 2026. Source for 1,000 agents on an early Claude Fable 5.2 with one lead, the Lean proof of Fermat's Last Theorem reproduced in 28.7 hours, and the September formalisation's 11 days. claude.dev
- [7]Anchor Terminal changelog, Claude Fable 5.2 and Claude Mythos 5.2 listed, 9 November 2026. Source for the release date and the prices, $8 input and $40 output per million tokens, down from $10 and $50 for the 5.1 models. anchorterminal.com
- [8]Koray Kavukcuoglu, Google, Introducing Gemini 4 Hydrogen, 12 November 2026. Source for Hydrogen going into Search for all users globally on launch day. blog.google
- [9]Meta AI Research, Introducing our smallest, most efficient model, Muse Ember, 9 November 2026. Source for Meta calling it its smallest and most efficient model. research.meta.ai
- [10]Meta, Pricing and rate limits, Meta Model API documentation. Source for $0.10 input and $0.50 output per million tokens. dev.meta.ai
- [11]Meta, Models, Meta Model API documentation. Source for Muse Ember being a hosted model with no open weights. dev.meta.ai
- [12]Amin Vahdat, Google, Our newest TPUs, 8tx and 8i Max, built for the agentic era, 9 November 2026. Source for TPU 8tx and TPU 8i Max being the third and fourth eighth-generation systems after TPU 8t and TPU 8i in April, and the first time Google introduced four TPU systems in a year. blog.google
- [13]Google Cloud, Tensor Processing Units (TPUs). Source for TPU 8t and TPU 8i listed as coming soon. cloud.google.com
- [14]Alibaba's Qwen team (@alibaba_qwen), post on X, 11 November 2026. Source for Qwen4-Max being in preview on Token Plan. x.com
- [15]Qwen Cloud, Token Plan individual, documentation. Source for the Lite plan at $6 a month and the list of models ending at qwen3.8-max. docs.qwencloud.com
- [16]Anthropic, Operate Claude Managed Agents inside your security perimeter, 12 November 2026. Source for self-hosting the whole service, the server, database and workers the customer runs, and the private beta. claude.com
- [17]Anchor Terminal changelog, Warnings on listings that look wrong, 13 November 2026. Source for the warning and its 15 points, its two kinds, and the 13 listings with nobody found behind them plus Pamba. anchorterminal.com
- [18]Anchor Terminal changelog, x402 counts for less (methodology v0.8), 13 November 2026. Source for machine payment protocols going from 40 to 10 of the 100 payment points and a tool with x402 and little else no longer outranking better ones. anchorterminal.com
- [19]Anchor Terminal changelog, the Terminal's filters stay light, 14 November 2026. Source for filtering no longer putting every listing's row in the page. anchorterminal.com
- [20]Anchor Terminal changelog, the people behind a listing, in more depth, 14 November 2026. Source for the eight people with a record of their own and every line linking its source. anchorterminal.com