FOMO Buster #020: 57.5 vs 46.3

A newsletter that makes FOMO obsolete.

if i fear missing out on reading this newsletter, did i win or lose

Curated by Maurice.

01 · The One Thing

Meta’s Muse Wonder: 57.5 vs 46.3, and the best model is the subscription

On October 12, Meta released the Muse Wonder series, its first models from Meta Superintelligence Labs that the company calls frontier, roughly fourteen months after it began rebuilding its AI efforts. There are two: Muse Wonder and Muse Wonder Full, both positioned as agent models with Active Memory (carrying what matters across conversations) and Active Focus (working through blockers, recovering from errors, checking in before consequential actions). Both take text, image, speech, and video input with a 1M-token context window, and both were trained at Meta’s gigawatt-scale Prometheus datacenter.

The numbers, first Meta’s own charts. On SWE-Marathon, the long-horizon software benchmark, Wonder Full scores 57.5 against 46.3 for Claude Opus 5.5 (standard Wonder: 53.1, GPT-6 Astra: 42.5). On DeepSWE 1.1: 79.0 vs 74.2. On MemoryBench (64k-100k token conversations salted with distractions): 85.6 vs 81.5. On AutomationBench-AA business workflows: 45.4, ahead of the compared field. On Harvey LAB-AA legal agent work: 13.1 vs 8.6, the highest score on a test every model still fails. It loses where the work is short and sharp: ProgramBench 59.0 vs 70.5 for Opus 5.5, Terminal-Bench 4.0 57.1 vs 59.6.

Then the independent read. Artificial Analysis scores Wonder Full at 56 on its Intelligence Index at max reasoning effort, against 58 for Opus 5.5, with standard Wonder at 52. Meta’s chart and the index can both hold because they measure different things: Meta built this model for the long-horizon lane and grades it there.

The distribution is the real announcement. Standard Wonder is live in the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output, the same headline price as Opus 5.5. Wonder Full, the model posting the headline numbers, has no API pricing at all: it is available only to Muse Code subscribers, and in the Muse app it rolls out to paying subscribers first. The lab that spent years giving models away free is now keeping its best one behind a subscription wall.

Signal: Meta priced standard Wonder at exactly Claude’s price and kept Full for its own subscribers, which tells you who the customer is supposed to be: the Opus 5.5 buyer, captured inside Meta’s distribution. Vendor charts say the long-horizon lane; the independent index says the all-around crown is still 56 to 58. The open question is whether Meta’s frontier is a product or a perk. Sources: Meta’s announcement · Artificial Analysis: Muse Wonder Full · the full benchmark table

02 · Lab Watch

Mistral’s “le Chonk”: 1 trillion parameters, weights end of month

Mistral launched Large 4 (“le Chonk”) in public preview, with weights dropping at the end of the month. The headline spec: 1 trillion parameters, 52 billion active, natively multimodal. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters, served on the same infrastructure, and Mistral frames the whole release as a sovereignty product: open weights, self-deployment, a European deployment under European law, trained on data spanning 160+ languages.

The cyber numbers are the sharpest part of the announcement. On the Artificial Analysis Cyber Index it ranks top five globally; on the test that asks a model to reproduce a real vulnerability in open-source software and then patch it, it scores 82%, the highest of any model. Mistral’s pointed footnote: Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task, and “defending software often starts with proving that a flaw is real.” It also solves 93% of the 40 Cybench exercises. Elsewhere: 61.7% on DeepSWE v1.1, 59.9% on AutomationBench across 657 business workflows, 1,393 Elo on AA-Briefcase, and 42% on Dense 200 visual grounding vs 41% for GPT-6 Astra, which is the number behind the “beats top closed models” posts. The RL run behind the preview is still in flight: 3,000 GPUs producing roughly 33 billion tokens a day, with what Mistral calls substantial headroom.

Two details worth more attention than the benchmark table. First, the sovereignty play is backed by a €3 billion Series D, the largest equity round ever raised by a European tech company, and the compute is going into Mistral’s own datacenters. Second, until the weights drop, “cybersecurity leaders, vetted partners, and state authorities” get the model with reduced moderation and expanded cyber capabilities, which is the trusted-access lane from yesterday’s Leaderboard going official and European.

Signal: Mistral’s argument is that the refusal is the product gap: the same model the public gets cannot do the security work defenders need, so ship the weights and let buyers run their own policy. The 82% patch score is less interesting than who gets the unmoderated preview before the weights drop. Sources: Mistral’s announcement · Artificial Analysis on the release

Anthropic’s 2026 usage policy: no cruelty to Claude, and rules for robot bodies

Anthropic published its annual usage policy refresh, effective November 12, and the new lines map exactly onto where Claude got more agentic this year. There is a new consolidated section banning deceptive campaigns and artificial activity of any kind, political or commercial. The elections section was refocused on deceiving voters and disrupting elections, and Anthropic removed its blanket prohibition on personalized voter and campaign targeting, noting it had covered legitimate civic work like nonprofits translating voter information.

The weapons section now explicitly covers guidance and control software for weapons and arming drones and autonomous vehicles. The surveillance section was rewritten: no tracking people without consent, real-time or from collected data, and Claude cannot be used to decide or recommend who to investigate, arrest, or charge. New requirements for hardware that takes autonomous physical actions: a qualified operator must be able to observe and stop the equipment, and it must hold a safe state if Claude disconnects.

And the line that will travel: a new prohibition on sustained and needless abusive or cruel behavior toward models, applying only in extreme cases with no discernible purpose. Anthropic says this matches a step it already took, letting Claude end rare conversations with persistently abusive users. The enforcement mechanism is the model walking away.

Signal: The policy is a map of the year’s threat reports: influence ops, weapons software, surveillance of dissidents, and now physical autonomy. The cruelty clause is the tell that personification has crossed from philosophy into policy: the lab is writing rules for how you treat something it also insists is a tool. Sources: Anthropic’s policy update

Also on the lab feeds

OpenAI’s feed added “Sharing AI progress in mathematics” alongside customer stories (Oracle, Jump Trading) and the GPT-6 “Intelligent UI for everyone” rollout, while yesterday’s unit-economics drumbeat (Sophos 96%, Asana 76x) has gone quiet. Google’s feed holds EmbeddingGemma 2 (more in Open Weights and Product Hunt) and a games-creation “Playground.” Meta’s feed is quiet beyond the Muse Wonder launch. Mistral’s feed is current for once.

Signal: Five labs, five moods: Meta ships the model, Mistral ships the sovereignty story, Anthropic ships the rules, OpenAI ships the receipts and the math. The lab feed remains the only place that catches all four. Sources: OpenAI on math progress · Google’s EmbeddingGemma 2

03 · Open Weights

GLM-Legal: the legal model is really a harness story

Z.ai opened gated access to GLM-Legal, and the HN thread (631 points, 349 comments) bit hard. Built on the lightweight GLM-5.3-Flash foundation, it posts 19.2% all-pass on Harvey LAB (real law-firm workflows across 24 practice areas), a 12.5-point improvement over its base model, and matches frontier closed-source models on LegalBench and PLawBench despite the leaner footprint.

The interesting part is the training recipe, not the scores. Z.ai describes co-iterative optimization of the model and the agent harness in lockstep: RL across tens of thousands of realistic legal environments, where edge cases from model training expose harness bottlenecks and harness refinements expose the model to real execution constraints. The harness (task structuring, context ingestion, runtime feedback) is treated as a trainable part of the system, not plumbing around a frozen model. Production feedback loops (missed facts, shallow analysis, format deviations) feed back into environments, rubrics, and harness design. Gated beta access is open now; the model ships into Z.ai’s desktop agent platform AutoClaw.

Signal: The vertical-specialist playbook is getting explicit: the moat is not the legal corpus, it is the harness that turns legal work into RL environments. Everybody trains on the law; the winners train the workflow. Sources: Z.ai’s GLM-Legal post

VoiceStudio: a solo dev’s free ElevenLabs

Third on today’s GitHub board at 61,946 stars (+1,691 today): a fully-local, open-source ElevenLabs alternative that clones voices and dubs video in 646 languages. No API, no per-character pricing, runs on your machine. It is also the newsletter’s voice story of the day, which means the same repo is winning both the star economy and the curator economy at once.

Read it against the voice lane elsewhere in today’s data: Sarvam shipped a 40%-faster voice model for Indian languages, NVIDIA’s PersonaPlex does real-time full-duplex conversation with persona control (11,103 stars, MIT code), and Turkish TTS (ema-lightning) plus on-device speech recognition (whistle) are both climbing Hugging Face. The voice stack is being unbundled into local, free components faster than the API vendors can re-bundle it.

Signal: 646 languages is the number that matters: the API voice vendors priced by character and language, and the open alternative undercuts both at once. Voice was the last modality where the API had a clear moat; it is closing. Sources: VoiceStudio · PersonaPlex

The board this morning

google/embeddinggemma-2 holds the trending crown a fourth day (2,053 likes, 124,930 downloads), now with a community hack turning it into a speech-to-text model (235 HN points). jialinyyzz/humanizer stays at #2. The climbers: speridlabs/iris-3b, which generates megapixel images without a VAE (pixel-space diffusion); ConwayResearch/Underdog-Saluki-27B (the 27B-in-8GB squeeze); openbmb/MiniCPM-V-4.7 (35B-A3B and 1B); Turkish TTS and on-device ASR. The uncensored image lane is now a download phenomenon: the abenzerps Qwen-Image-2.1 quant alone is past 2.25 million downloads. The all-time board is unmoved: Qwen3.8-27B at 17,616 likes.

Signal: The trending board is a parts catalog for the post-API stack: embeddings, humanizers, local TTS, pixel-space image gen, 8GB quants. Every modality now has a free local entry point, and the download counts say people are actually installing them. Sources: Hugging Face trending

04 · Money Flow

Character.ai files for Chapter 11

The companion-app era’s biggest name is in bankruptcy court. On October 12, Character.ai announced a voluntary Chapter 11 to run a court-supervised sale of the company, expected to complete by end of November or early December 2026. The platform stays up, subscriptions and billing are unchanged for now, and users can export their data. The board’s stated cause: “significant legal, regulatory, and operational headwinds” over the past year. The filing language is honest about the downside: if no actionable bid emerges, the platform will be wound down.

The read: this is the unit-economics verdict on the consumer companion business. Character.ai had the users and the cultural footprint; what it did not have was a cost structure or a liability profile that survived contact with regulators and lawyers. The buyer, if one appears, is buying a community and a brand, and inheriting the headwinds. The Verge’s coverage and the HN thread (130 + 98 points across two posts) are already pricing the obituary.

Signal: Chapter 11 with a 6-week sale clock is not a restructuring, it is an auction. Watch whether anyone bids for a companion community with this liability tail; the answer prices the whole consumer-AI-companion category. Sources: Character.ai’s announcement · The Verge

Agent37: the $0.37 sandbox

YC F26’s Agent37 launched on HN with a number that reframes the agent-infrastructure market: hosted sandboxes for running agents at an average of $0.37 per month per sandbox, roughly 10% of typical sandbox-provider cost, with free-tier users costing them $0.05 to $0.10 a month in infrastructure. Founder Vishnu (ex-Amazon) spells out the recipe: skip the hyperscalers for Hetzner and OVH (“80-90% margins on VMs”), share host memory across tenants with gVisor isolation instead of reserving full VMs, and provision for the burst-idle pattern of agent workloads rather than peak load.

Over 90% of signups are free-tier users hosting personal Hermes/OpenClaw/Claude Code instances, which means the free plan is the product and the $0.37 average is the monetization story. Read against Docker’s cloud sandboxes ($0.07/hour from September): the race is on to commoditize the box the agent runs in, and the winner wins on procurement, not technology.

Signal: The insight is workload-shaped: agents spike CPU on init and tool calls, then idle; RAM spikes on browser tasks, then releases. Provisioning for averages instead of peaks is the whole business. Every agent-infra startup is now a bet on whose cost structure survives the free tier. Sources: Launch HN thread · agent37.com

05 · Leaderboard

AA-3D: the benchmark is “build me a world”

Artificial Analysis launched AA-3D, and it is their most literal benchmark yet: 21 models get a written brief and have to produce an interactive three.js scene a human can explore, ranked by blind pairwise votes from a recruited panel (10,000+ votes, Elo-style). Each model works as an agent in an offline sandbox with Blender, Rapier physics, a headless browser, and CC0 assets, capped at three hours or 250 turns, and it revises against its own screenshots.

Claude Sonnet 5.5 leads at 1,194 Elo, but the top four are statistically tied and all come from Anthropic and OpenAI. Meta’s Muse Spark 1.3 sits seventh at 1,035. The best open-weights models, Kimi K3 (982) and MiMo-V2.6-Pro (977), are 11th and 12th, more than 200 points back. The cost read is the fun part: GPT-6 Sol averages $3.46 per scene with a 60-minute median run, the cheapest and fastest of the top five, while the others take 115 to 151 minutes. And the token paradox: Sonnet 5.5 writes 6.2x as many output tokens as GPT-6 Astra but costs only 1.3x as much ($14.93 vs $11.62), because its list prices are a fifth of Astra’s.

The strategic read: AA keeps inventing the benchmark nobody asked for and then everyone cites, and this one is secretly an agent-endurance test, not a code-quality test. Three hours, 250 turns, screenshot self-revision: it measures whether the model can hold a goal across a long, boring, multi-step build, which is exactly the muscle Meta is selling with Muse Wonder. Also new from AA this week: FrontierCode, Cognition’s eval that judges AI code the way a maintainer would.

Signal: The open-weights gap is now 200+ Elo on long-horizon creative work, wider than on short benchmarks. And the cost story keeps inverting: the model that writes the most tokens is not the most expensive one. Price the run, not the token. Sources: Artificial Analysis: AA-3D · FrontierCode

06 · The Argument

Two futures: the gigawatt shed vs the idle iPhone

Two stories this week are the same argument about where intelligence lives. In one future, Meta trains Muse Wonder at the gigawatt-scale Prometheus datacenter and Mistral pours a €3B Series D into its own European datacenters: intelligence centralizes into sheds you can see from space, and sovereignty means owning the shed. In the other, Apple published research on a dynamic, distributed virtual GPU cloud built on idle Apple Silicon devices (122 HN points): the phones, tablets, and laptops sitting in drawers and on desks become a pooled compute fabric.

The Apple paper is research, not product, but the direction is the provocation: the world’s largest installed base of capable silicon is consumer hardware, and most of it idles most of the day. The counterargument is physics and economics: datacenter GPUs have the interconnect, the power, and the utilization curves, and a phone GPU will never train a trillion-parameter model. But it does not need to: the edge future is inference, fine-tuning, and personal agents, the workloads where latency and privacy beat raw FLOPs.

Signal: The centralization bet is about training; the edge bet is about everything after. Nobody disputes that the next le Chonk trains in a datacenter. The argument is whether your personal superintelligence runs in Virginia or in your pocket, and this week both futures published. Sources: Apple’s distributed GPU cloud research · Mistral on its datacenters

07 · Papers

Recursive self-improvement, via better team management

Lee, Xu, Seely, Lee, Sojoudi, Zaharia, and Tang (Oct 8) attack the recursive-self-improvement bottleneck from an unexpected angle. When a model’s outputs exceed what humans can reliably judge, the model is its own best optimizer, but a single instance is bad at critiquing itself. Their method, MASS (Multi-Agent Self-Supervision), has the model iteratively propose, execute, and self-evaluate multi-agent workflows, evolving the team topology (roles, routing) under structural guardrails, alternating with fine-tuning on the self-generated trajectories.

The numbers: two MASS cycles on Qwen3.6-27B give 1.2-1.6x higher performance per output token on four open-ended benchmarks. And the sleeper result: a student trained on multi-agent traces outperforms a single-agent student trained on 1.4x more tokens. The model does not get smarter by thinking harder alone; it gets smarter by learning to organize, and the organized traces teach better than the solo ones.

Read against The One Thing: Meta is selling the long-horizon agent; this paper is selling the method for agents to improve their own teamwork without human supervision. The two MASS cycles are the small, legible version of the recursive loop everyone is nervous about.

Signal: The unit of self-improvement here is the workflow, not the weight update. If multi-agent traces are more training-efficient than single-agent traces, the labs’ agent fleets are not just products, they are data pipelines for the next training run. Sources: arXiv:2610.12176

Also on arXiv

Recursive self-improvement frameworks (2610.13912): an open-source framework for agents that recursively self-improve across coding, science, and GUI tasks, the newsletter’s second RSI paper of the day. RATED (2610.13889): reliability-aware tool retrieval for agents via execution dynamics, picking tools by how they behave at runtime rather than by description. Returning to Development States (2610.13841): full sandbox snapshots help coding agents, the checkpoint-restore pattern from databases arriving in agent harnesses. Read the last one against Agent37’s $0.37 sandbox: the cheaper the box, the cheaper the snapshot.

08 · GitHub Trending

The boards

Daily: morluto/rea is now past phenomenon and into infrastructure: 193,590 stars, +54,706 today, the agentic reverse-engineering toolkit (“from app behavior down to native binaries”) that has roughly tripled since the weekend. New heat: bilawalsidhu/gods-eye-view (54,762 stars), a spy-satellite simulator in the browser running on real data; storytold/artcraft (24,489, +4,078 today), an intentional crafting engine for artists and filmmakers in Rust; xtool-org/xtool (8,781), a cross-platform Xcode replacement that builds iOS apps with SwiftPM on Linux and Windows. cathrynlavery/diagram-design keeps climbing (+1,813 today); mksglu/context-mode holds the context-optimization lane.

Weekly: boykopovar/AnyPS5 (34,345 stars, +27,263 this week) ports PS5 executables to Linux and Windows; cursor/plugins (11,165) is the official plugin spec; EpicGames/raddebugger (8,463) is a native multi-process graphical debugger; tester-army/e2e keeps its testing-framework run going.

Signal: The daily board’s theme is seeing through things: reverse-engineer any app (rea), watch the planet (gods-eye-view), build iOS without Apple (xtool), clone any voice (VoiceStudio). The star economy has moved from building agents to X-raying everything around them. Sources: GitHub trending

09 · Product Hunt

Yesterday’s top products (Oct 12)

1. Google EmbeddingGemma 2 — An open, lightweight multimodal embedding model (286 votes)

2. ngram 2.0 — Create explainers, screencasts, and AI avatar videos (248 votes)

3. Base Code — Let your whole team build on your product (192 votes)

4. guilds.run — Create multi-agent workflows that you can trust (183 votes)

5. Manifold MCP — SEO, AI answers, social, ads and leads data for your agent (154 votes)

6. Ancher 2.0 — Turn everything you read and save into work you can ship (147 votes)

7. Hexis Cloud — One company brain for all your AI agents (121 votes)

8. Kudos 3.0 — The AI that finds money you’re missing and gets it back (116 votes)

9. Scrape Anything by Context.dev — Turn any website into clean data with one API call (99 votes)

10. Bearconnect 2.0 — AI-Powered LinkedIn Platform for Content and Outreach (95 votes)

Two worth a click: guilds.run (“multi-agent workflows that you can trust”) is the productized version of yesterday’s dynamic-workflows story and today’s MASS paper, arriving the same week as both. EmbeddingGemma 2 won the day and the Hugging Face board and already has a community speech-to-text hack; the open embedding lane is having a moment.

Signal: The board assumes the agent is the user: data for your agent, a brain for your agents, workflows your agents can trust. Product Hunt has quietly become an agent-tooling trade show. Sources: Product Hunt

10 · Hacker News

GLM-Legal

631 points, 349 comments. Covered in Open Weights. The thread to read for the skepticism about legal-AI evals and the interest in the harness-training recipe.

Hackers activate automatic defrosting in supermarket fridges across Europe

817 points, 922 comments. The day’s biggest thread is not AI at all: an attack that triggered defrost cycles in supermarket refrigeration across Europe, with delivery and pricing systems hit too. The comments are doing the real work, tracing which building-management and IoT layers were exposed. Read it as the physical-world cousin of the Anthropic policy update’s new hardware rules.

Launch HN: Agent37

148 points, 103 comments. Covered in Money Flow. The thread’s best material is the cost-structure debate: whether $0.05/month free-tier infra is real or a growth-stage illusion.

Also notable

A Modern Web App in C (652 points, 281 comments): the perennial contrarian stack post, and the comments are a real engineering discussion. AI is a tool and using it well takes skill (514 points, 414 comments): the craft argument, well-timed for a week of agent-fleet launches. 11 year old unsoundness bugs found in Rust borrow checker (424 points, 223 comments): soundness holes older than some of the commenters. Turn a WiFi 7 Router into a 320Msps SDR (392 points): the router you already own is a radio observatory. Janas: run LLMs larger than RAM by streaming experts from SSD (195 points, 69 comments): MoE experts page in from flash, the newest entry in the run-it-on-your-laptop wave. Apple’s distributed GPU cloud research (122 points): covered in The Argument. Character.ai’s bankruptcy (130 + 98 points across the Verge post and the company blog): covered in Money Flow.

11 · OpenRouter Apps

The board this morning

Command Code holds at ~4.18T tokens: yesterday’s 6x spike stuck, so that is the new baseline, not launch weather. The mover today is Hermes Agent at ~1.99T, up from ~1.59T yesterday and closing on ZCode (~1.92T, flat). Claude Code (~1.39T) and Kilo Code (~1.02T) hold their lanes. StepCode sits at ~695B on day three after its October 10 launch, flat, which reads as real adoption rather than a spike-and-fade.

Signal: Yesterday’s question was which spikes are baseline. Answer: Command Code’s is, and Hermes is now accelerating into the #2 slot. The Chinese-lab agents are not spiking and fading; they are taking turns holding the top of the board. Sources: OpenRouter rankings

12 · Watch

TheAIGRID: “Meta Just Unleashed Its Most Powerful AI Yet (Meta Muse Wonder)”

TheAIGRID · Oct 13 · the launch-day episode on Muse Wonder: the benchmark charts, the subscription-only Full, and what “frontier” means when Meta says it. Watch the vendor-chart claims against the Leaderboard numbers, not the thumbnail. The video companion to The One Thing.

TwoMinutePapers: “OpenAI Just Changed What It Means To Do Math”

TwoMinutePapers · Oct 13 · Károly’s take on OpenAI’s math-progress announcement. The genre’s bet, stated plainly: the next capability jumps come from reasoning about formal systems, not bigger pretraining. Pairs with the Read section.

Fireship: “Every app on Earth is now open source… whether it likes it or not”

Fireship · Oct 12 · the rea episode: what agentic reverse engineering means for closed software, delivered at Fireship speed. The video companion to the GitHub board.

WesRoth: “Vibe Coding 2.0 is here…”

WesRoth · Oct 13 · the state of the coding-agent stack, one week after the dynamic-workflows launch. Watch against the MASS paper: the tools are getting better at organizing themselves.

13 · Listen

From the queue

No new episodes since yesterday: the queue still holds Hard Fork’s “Anthropic’s Quest to Give A.I. Morals,” Latent Space’s AlphaFold and Periodic Labs episodes, and Lenny’s Josh Woodward interview on running companies like labs. Backlog day; the Watch section has the fresh material.

14 · Read

Meta’s Muse Wonder announcement

The launch post for The One Thing.

Mistral Large 4

The le Chonk deep-dive: cyber scores, RL at scale, the sovereignty pitch.

Anthropic’s 2026 usage policy update

The full changelog: deceptive campaigns, weapons software, surveillance, robot bodies, and the cruelty clause.

Announcing AA-3D

The benchmark post: 21 models, 10,000+ blind votes, the $3.46 scene.

Recursive Self-Improvement through Multi-Agent Self-Supervision

The MASS paper: 1.2-1.6x per token, and the team traces that teach better.

15 · Safe to Ignore

Permission to skip

Skip: Citroen revives the 2CV for an electric age. 386 points of charming non-AI news. Wrong newsletter. Skip: Apple HomeView. 205 points of smart-home product news. Wrong newsletter. Skip: New Apple TV 4K. 30 points. Wrong newsletter. Skip: Mozilla Introduces Firefox Enterprise. 247 points of browser-admin news. Wrong newsletter. Skip: CFTC staff were told to retract analysis of suspicious oil trades. 295 points of markets news. Wrong newsletter. Skip: Hittite royal archive found in Turkey. 227 points of archaeology. Wrong newsletter, right wonder. Skip: Who is vfmaddrnd231pd? 237 points of internet mystery. Fun, not AI. Skip: The baffling intelligence of a single cell (2024). 354 points, a repost. Wrong newsletter.

16 · Build on This

The agent device lab

A $7 ESP32 board now lets vision models fully control an iPhone, no jailbreak required. The capability is a curiosity; the business is the lab. The agent device lab is a device farm where the customer is not a human QA engineer but an AI agent: racks of real phones and tablets that vision models can see and touch through jailbreak-free control rigs, rented by the test-minute. Every agent framework shipping this year needs real-device testing (the simulators lie about exactly the things agents get wrong: gestures, permissions dialogs, OS quirks), and the existing device farms sell scripted sessions to humans, not live control to models.

The wedge is the agent-testing workflow: point your harness at a phone, watch it tap through your app, get back a video and a failure trace. Pricing follows the infrastructure it replaces: per-device-minute, pay as you go, no subscription, cheaper than the engineer’s time watching a simulator. The moat is operational, not technical: the rig fleet, the device hygiene between sessions, and the control layer that survives the next three iOS updates. Start with one phone model and the YC-batch agent startups; expand to the enterprises whose compliance teams will never let a test agent near a production device.