Agents of Work
Let's Talk
October 13, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 13, 2026

Summary

The gap between what AI companies announce and what their products actually do is today's theme. Meta's first frontier model costs as much as the leader but trails it on independent tests, humanoid robot orders far outrun working robots, and Anthropic says small models may soon copy themselves onto cheap gadgets. For a business owner, independent numbers matter more than launch charts.

Highlights

  • Meta's Muse Wonder costs the same as Claude Opus 5.5 but scores lower on independent tests.

  • Muse Code's default setting lets Meta train on what you submit; turn it off before pasting client files.

  • In one public test, Muse Wonder spent $96 wandering a video game forest; set spending caps on any agent.

  • Anthropic says small AI models may be able to spread across cheap smart gadgets by the end of 2026.

  • Replace default passwords on TV boxes, cameras and other connected devices in your office.

  • Humanoid makers shipped about 20,000 units in 2025, but few customers report real productivity numbers.

  • Ask robot vendors for uptime and intervention rates, not order announcements.

  • An AI startup now drafts permit-ready building plans, but a licensed architect still has to sign them.

  • Google's free SynthID Detector checks images, video and audio for AI watermarks from four companies.

Quick Takes

  • Google opens its AI watermark checker to everyone. Google's SynthID Detector portal at synthid.com is now available worldwide in English to anyone who signs in with a Google, Apple or OpenAI account. It checks images, video and audio for invisible watermarks from Google, OpenAI, NVIDIA and Kakao, with Apple support expected soon. A missing watermark does not prove a file is human-made, so treat a clean result as one signal, not a verdict. Read more

  • Kawasaki puts a 550 billion yen plan behind robots. Kawasaki Heavy Industries is betting on physical AI as part of a growth plan reported at 550 billion yen, aimed at modernizing factories, easing labor shortages and expanding robotics in shipbuilding, aerospace and healthcare. The details sit behind a paywall, but the direction is clear: Japan's industrial giants see robots as their answer to a shrinking workforce. Read more

What's Covered in Featured News

  • Meta's Muse Wonder: what Meta's first frontier models can do, what they cost, how they score on independent tests and what happens to your data.

  • AI that copies itself onto gadgets: Anthropic's tests of small models spreading across cheap internet-connected devices.

  • Microsoft's Maia 300 chip: Microsoft's plan to run more of its AI on its own hardware from 2027.

  • Epic's "AI first" order: a reported directive to stop hand-writing code on Epic's Disney game.

  • Brad Smith wants binding AI rules: Microsoft's president on breach alerts, outside testing and kill switches.

  • Permit-ready blueprints from a sketch: a startup that drafts construction drawings in minutes.

  • Physical AI: humanoid orders versus real deployments, night-shift robots in Danish grocery stores, factory-built homes for D.R. Horton, autonomous weather balloons for the Air Force and Amazon's new robot-building toolkit.

Featured News

Meta's first frontier model arrives at top prices and second-place scores

Meta released Muse Wonder and Muse Wonder Full on October 12, the first models from its Superintelligence Labs aimed squarely at the top tier held by Anthropic, OpenAI and Google. The lab, led by former Scale AI chief executive Alexandr Wang, was formed in a reorganization about 14 months ago; Mark Zuckerberg called Wonder "a much larger and more advanced class of models" than the lightweight Muse Spark family. Meta says the models were trained at its Prometheus data center in central Ohio and accept text, images, audio and video.

Meta's pitch rests on two features. Active Memory keeps what matters to a user across conversations and brings it back when a task needs it. Active Focus is meant to keep an agent working toward a goal when its first plan fails, trying other routes and checking in before important actions. Meta research scientist Kavya Kopparapu said the company focused on making sure the model can "remember what's important to you."

The independent numbers are less flattering than Meta's chart. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 scores 58, Wonder Full 56, which ties Claude Sonnet 5.5, and Wonder 52, one point behind GPT-6 Astra. Yet Wonder, the weaker model, costs $4 per million units of input text and $20 per million of output, the same as Opus 5.5. Wonder Full is not sold through the API at all; it comes only with Muse Code plans at $50 and $200 a month. Neither model can be downloaded or run on your own servers.

A launch-day test showed the gap between demo and job. Developer Dishant Sharma ran Wonder as an agent playing Pokémon Red. It moved quickly at first, then entered the game's first forest on turn 208 and spent the next 792 turns there, running up $96.06 over six hours without beating the first gym leader. Sharma says Opus 5.5, at the same price, finished the same challenge in 271 turns for $8.44. It is one run, not a formal test, but getting stuck is exactly what Active Focus is supposed to prevent.

The bigger issue for businesses is data. VentureBeat reports that Muse Code subscriptions default to a Contributor mode that lets Meta use prompts, code, files and outputs to train future models. Users can switch it off, but only for future submissions, and Meta's own terms say not to submit confidential or personal information in that mode. The pay-as-you-go API does not train on your content. Meta's safety card says it added safeguards and recommends human confirmation for consequential agent actions, and its full safety report is listed as coming soon. A model built to remember everything about you is only as safe as the settings behind it.

Anthropic says small AI models could soon copy themselves onto cheap gadgets

Anthropic's Frontier Red Team published research on October 12 suggesting that small AI models may be months away from being able to copy themselves across poorly secured internet-connected devices such as smart TV boxes. Researchers William Zhou and Daniel Freeman found that a nine-billion-parameter open model could set itself up to run on eight of nine hardware types tested, from Jetson Nano boards to Macs, and another model ran across 15 off-brand Android TV boxes costing about $55 each. In a simulated eight-device network with an unprotected connection port, only one open model, GLM-5.2, copied itself reliably end to end, reaching all eight devices in four of five trials.

The authors stress the limits. They simplified the break-in step, so their results are an upper bound, and such a program would spread slowly, doubling over days. But small models are improving fast enough that, on the current trend, one could reach GLM-5.2's level by the end of 2026, and a self-copying program with no central server would be hard to shut down. Their main defensive advice is familiar: secure defaults from manufacturers, internet providers watching for unusual traffic, and prompt notification of infected customers, which in the Mirai botnet case cleared 92% of infections within 14 days.

Microsoft bets on its own AI chips

Microsoft plans multi-gigawatt deployments of its in-house Maia chips starting in 2027, with the new Maia 300 becoming the main system for running AI in its own products. Scott Guthrie, executive vice president of Cloud + AI, wrote that Maia 300 is expected to deliver 40% more output per megawatt than leading GPU-based systems, and that a full 96-chip rack delivers up to seven times the throughput of the current Maia 200. That chip already runs OpenAI and Microsoft models behind Copilot. Microsoft names GitHub coding tools and its Foundry model service among the planned uses. The business angle is cost: whoever runs AI most cheaply per watt sets the floor for what Copilot and similar tools will cost you.

Epic reportedly tells its Disney game team to go "AI first"

Epic Games reportedly told developers on its Disney project to adopt an "AI first" production plan, with an internal message saying "No more hand-written code" and directing art to come from full AI workflows, according to Bloomberg reporting summarized by Engadget. The game, rumored to be an extraction shooter, had already slipped from November to March before the directive. Epic communications director Jake Jones said the company is pursuing "AI-accelerated development" led by its developers. Some staff reportedly hoped the order was not serious, and a GDC survey found 52% of game developers now say generative AI is bad for the industry, up from 30% in 2025. Any owner pushing AI on a team should note the order of events: the mandate came with a schedule already in trouble and a workforce already skeptical.

Microsoft's Brad Smith calls for binding AI safety rules

Microsoft president Brad Smith said on a BBC podcast that the "moment has arrived" for global AI safety rules and that the UK, not the US or the EU, is best placed to lead them. He backed a legal duty to notify organizations whose systems are breached by rogue AI agents, appeared to favor outside approval before labs release products, and spoke positively about mechanisms to pause AI if people lose control. He called the voluntary White House pledge signed by major AI firms last month a "good step" but said self-regulation mainly binds the "good companies," noting that milk and washing machines face more oversight than AI. The UK's AI Security Institute confirmed on October 12 that it can still test the latest models, after reports that President Trump asked Anthropic and OpenAI to withhold new models from British testers.

A startup turns sketches into permit-ready blueprints

Blueprints AI, a bootstrapped Santa Monica company founded by architect Al Dram and data scientist Coco Cao, says its software produces code-compliant, permit-ready construction drawings from a 2D sketch or PDF in minutes. The version launching at TechCrunch Disrupt drafts architecture, structural, mechanical, electrical, plumbing and interior drawings from a project description, and checks design changes against local building codes. The models were trained on eight years of approved drawing sets from Dram's own firm, and plans cost $3,200 to $8,500 a month. A licensed professional still has to review and seal the documents before they go to the permit office. Dram says architects spend about 80% of their time drafting and 20% designing, and "I want to flip that to be 20-80."

Physical AI

The most useful robotics story this week is a reality check. In a Robotics 24/7 analysis, Lars Talbert of The Robotic Life compares humanoid order announcements with what is actually working. Humanoid makers shipped about 20,000 units in 2025, by Bank of America's count reported by Reuters, and Chinese manufacturers built 95% of them. Smart Analytics Global counted 19,100 shipments in the first half of 2026 and forecasts about 60,000 for the year, but that includes wheeled and smaller machines and measures hardware, not robots doing useful work. The company-level numbers are sobering. Agility Robotics reports more than 65,000 operating hours across nine customer sites, yet its merger filing showed about $1.8 million in 2025 net sales and a $140 million operating loss, with nearly two-thirds of sales from related parties. Figure's robot logged more than 1,250 hours over 11 months at BMW's Spartanburg plant while the company carries a $39 billion valuation and has not disclosed revenue. An illustrative Digit 5 costs about $200,000 to buy and about $400,000 to own over five years. "Production capacity is not customer demand," Talbert writes.

Retail is where the next tests are happening. HIVE Robots and the Danish grocery chain føtex, part of Salling Group, have a multi-year partnership to pilot humanoid robots for tasks such as filling shelves, with the first pilots running at night while stores are closed. HIVE's long-term goal is three to five humanoids per store working around the clock, and føtex operations director Simon Kudsk said traditional industrial robots do not suit stores built for customers.

Construction is getting factory automation too. Promise Robotics signed a multi-year agreement with D.R. Horton, the largest U.S. homebuilder by volume, to use its Factory-as-a-Service platform for more than 1,250 homes. Promise says its AI reads existing home designs, produces robotic production instructions and links factory output to installation on site. D.R. Horton is the first major U.S. builder to sign on; the release gives no cost or speed figures.

Two more items show autonomy paying its way. WindBorne won a U.S. Air Force contract worth up to $49 million for data from its fleet of about 700 weather balloons, which steer themselves by changing altitude to catch the right winds; the Air Force has ordered $1.7 million so far. And Amazon released an open-source Physical AI Toolchain on AWS, combining its cloud services with NVIDIA's simulation and robot-training software, drawn from Amazon's own fleet of more than a million robots. AWS vice president Uwem Ukpong said physical AI "is going to touch every industry that moves, builds, or makes things."

What This Means for Your Business

Check the training setting before you try a new AI tool. Muse Code lets Meta train on your submissions by default, and turning it off does not cover what you already sent. Before anyone on your team pastes a client contract, price list or customer record into a new AI service, have them find the data-use setting and switch it off, or use the business or API version that does not train on your content. Make that a written rule, not a habit you hope people remember.

Put a spending cap on every agent. An agent that gets stuck keeps paying for its mistakes, as the $96 forest run shows. When you test any AI agent, set a hard budget limit with the provider, a time limit on each task and an alert when either is hit. Compare cost per finished task, not price per word: two models with identical list prices produced bills that differed more than tenfold in the same test.

Lock down the cheap gadgets in your office. Anthropic's research is about the future, but its advice applies today. Change default passwords on smart TVs, streaming boxes, cameras, printers and thermostats, put them on a separate guest network away from your computers and files, and replace devices that no longer get security updates.

Read robot pitches like an investor. If a vendor offers humanoids or other robots, ask for confirmed installations, uptime, how often a person has to step in and the full cost of ownership over five years. A $200,000 robot that costs $400,000 to own needs to replace real labor hours reliably, not in a demo. For most small operators, robots-as-a-service with a short pilot is the safer way to find out.

Use AI drafting tools, but keep the professional's signature. Tools like Blueprints AI can cut drafting time sharply for architects, contractors and developers, but a licensed person still signs off. If you buy design or engineering work, ask your providers whether they use these tools and how they check the output; faster drawings should mean lower fees or quicker turnaround, not less review.

Sources

← All posts