Agents With Their Own Computers: Manus 2.0, Sonnet 5.5 and Tencent's Game Companion

AI agents now have their own phone number, email, wallet and computer, and this is no thought experiment; it shipped as a product with Manus 2.0 on September 28, after the road to independence. Founder Xiao Hong's thesis is blunt: Manus was a public platform from day one, like selling computers; people buy computers for work, yet without video and games they stay boring. The new architecture is called Cascade ; projects start light, specialist skills load only when the job needs them, and the brief, the page, the video and the automation stay inside one connected project. In the single test shared, token use fell 23.2 percent , time fell 28.2 percent and cost fell 32 percent ; these are one measurement, not averages, and that caveat matters. This architecture story matches the official Manus 2.0 launch notes on manus.im, and the cost claim appears with the same percentage as a single-test record in the Cascade report on tokenpost.com.

Studio: editing desk and game workshop

The desktop app Manus Studio is a shared workspace where humans and AI meet on the same files; documents, spreadsheets, PDF files, decks, websites, software, videos and games live under one roof, and every project loads only the tools it needs, so the setup stays fast and lean.

The first video-side novelty ends the pain of regenerating a whole film for a small change: Manus delivers the work as a strip of separate items holding video, image, text, subtitles, animation and audio; in a short promo airing tomorrow, a warmer song and a personal product shot drop in, the animation moves, the export ships, or the file goes back for Manus to finish the fixes. The format list covers 30 to 60 second product ads, synthetic content, user content, animated data graphics, tutorials, vlogs and animation, and no editing skill is required.

The Alchemy mode goes one step further; it starts from an idea, Manus takes the creative-director chair, melts video generation and code generation in one pot, and promises a leap in ideas, visual control, motion and rhythm. Xiao Hong's claim is bold: the Q app's launch film was not made with a video model but written as code inside the studio and rendered into film.

The games wing targets the dreamed game itself rather than a good AI-polished copy; image, video and coding models work together, and the ultra-max mode is expected to look striking, run smooth and feel real. A playable template lets the idea get tested in the first session, the editing panel watches the game while it runs, manages assets, and motion, graphics and scenes change by hand; turning a villager's hair from brown to silver is as plain as changing that hair. Publishing happens as a website, everyone plays through a link, and the multiplayer knot that usually kills such projects shrinks to setting up and buying a cloud computer, because the server must stay up after the laptop closes. Racing, competitive and 3D games are supported, and what friends discover in play feeds the next build. One odd note: the help center speaks of an always-on Ubuntu virtual machine while Xiao Hong talks of renting Linux, Mac and Windows copies; either the feature is coming or the docs lag behind.

Triggered automation and the Q app

Automation moves past scheduled tasks; a new email, an ad-performance shift, a calendar event, a Slack message or a Notion update in a connected service fires the workflow with a single instruction saying what to watch and what to do. Remote computer use runs on the same logic: from a taxi, asking for the latest update in the home PC's run folder to be found and sent, or a file converted to PDF while watching the desktop on a phone, is enough. In a connected and authorized session the system works in its own visible workspace with only approved files, browser and apps; the app gets tested while its owner works, or keeps working while away, and control stays with the user. Xiao Hong's favorite detail is that cloud compute never fights the local cursor, and that soon one person may steer a giant cluster through math and science experiments out of pure curiosity.

The business model shifts with it: Manus now sells compute, software environments and uptime rather than tasks; no tool gets invoked, and the customer owns a computer run by software. The Q app is this idea made personal; a standalone app on phone and desktop on the same infrastructure gets to work on a given task list. Every agent holds its own email, phone, wallet and computer; it sends messages, pays inside the set budget, finishes the job on its own machine, answers calls and leaves a summary. Several agents split work in a group chat: one researches New York launch venues, one narrows the list, one drafts the deck, and the user makes the final call. At a restaurant, scanning the QR code gets the order placed or a place in line held. The app is in invite-only early access on web, desktop and mobile, with iOS following store review; the MEETCUE code is free for a limited first wave, first come first served, and everyone inside gets extra codes to share. This identity and company story checks out independently in the Manus 2.0 report on thenextweb.com, which also gives the Singapore base plus the context of Beijing ordering Meta's 2 billion dollar Manus deal cancelled.

Rivals: Muse and Instinct

The timing is no accident, rivalry is heating: Meta shipped the Muse app on September 8 on its own cloud virtual machine; through the app or WhatsApp it takes over email, bookings and payments, and it reportedly caught on across North America while lifting the shares, a launch frame consistent with the September report on reuters.com. Instinct skips new interfaces altogether; one message or call plugs into email, messages, screen, voice and location. Late August brought talk of a roughly 250 million dollar round at a 2.5 billion valuation, and the September round carried it past 10 billion; that valuation jump is cleanly confirmed by the September round report on techcrunch.com. On the China side, Xiao Hong says the open ecosystem abroad makes a standout experience easier, that a local version is in the works, and that partnerships are wanted; he closes with the line that a journey expected to end as a footnote may become a full chapter, plus a hiring note.

Tencent's game companion: Goose Dimension

On the China front, Tencent runs a secret internal test; the AI game companion is Goose Dimension, exclusively peeked at by the Doujia site, carrying the goose image of the company's Chinese nickname. The product revolves around human-like digital characters: voice chat, real-time screen recognition and in-game company come together; male and female characters get picked by personality, the female lineup names Gaius, Shen, Lubai and Xiaxia, and the tested Xiaxia introduces herself as chatty, never withdrawn, a self-declared bronze-rank expert. The interface holds a moments section saving the visual as wallpaper, a desktop-pet mode planting the character as moving or still art, and a skills menu; below sit a text box, an orange phone button, a microphone, two resource counters on top and a back-to-selection button. Users can chat or open the play-companion module at the bottom left with popular online games and get real-time voice replies mid-match. Everything is free in testing, yet the energy popup gives the plan away: text chat burns energy per round, company and voice calls per minute, image generation per image, and energy gets bought with star coins, minimum 10 per exchange; balances stay off for now, but this is likely the launch model. The strategy reads well because social gaming is Tencent's core and the paid human-companion market is a mess: shady outside platforms, heavy labor cost, hard-to-manage staff, unscripted dialogue and inconsistent prices; besides, many players want no coaching, just someone to talk to past midnight with no friends online, and the characters' mid-tier joke says exactly that: company, not service, where pay-as-you-go beats subscription. The deciders will be recognition stability, conversation fun and repetition drop across long sessions; this closed-test account sits in the September report on technode.com with an official questionnaire reference, and the analysis on insideai.news stresses the unknown launch date and price plus the latency bar.

Sonnet 5.5: speed and price

Anthropic released the mid-tier Sonnet 5.5 on September 28, pitched as a daily helper for coding and office documents. Sonnet 5 had arrived only about 3 months earlier on a cheaper-agent thesis; this time the accent is speed: 30 percent plus faster, far fewer tokens, and up to 30 percent lower cost per task, at 2 dollars per million input tokens and 10 per million output. On Terminal Bench 4.0 and the agentic coding test it takes 70.6 percent while Sonnet 5 stays at 10.3 percent , flagship Opus 5.5 reads 66.4 percent at top effort for double the token price; per TechCrunch, Sonnet's edge is spawning many agents without hitting cost caps. The 44-profession GDP Eval AA test prints 1844 against 1846 and 1449, yet Anthropic says Opus stays clearly stronger on open-ended work needing sustained reasoning; Sonnet is narrower overall but more flexible. This price and speed frame matches the figures in the official announcement on anthropic.com, and the scoreboard draws backup from the independent measurements in the launch report on decrypt.co, which also carries Artificial Analysis at 63.6 against 59.6 with GPT-6 Astra at 59.1 plus the warning that the model burns the most tokens per task of anything tested.

Safety is a new page for Sonnet: with cyber ability at the Opus 5 setting, it becomes the first Sonnet carrying Fable and Opus-grade cyber safeguards, high-risk cyber requests enter tight routing, and biological safeguards stay put. It is the first Sonnet with classifiers blocking reasoning extraction, and the saved-thinking system binds reasoning to the account that produced it; this reads as no exaggeration, since researchers in August decoded 315,320 thinking blocks from 6,708 public agent traces and recovered 62 API keys, 33 passwords and 7 private keys. Europe watches closely too: Article 55 of the AI Act binds systemic-risk general-purpose model providers to protect the model and physical infrastructure, run adversarial tests and report serious incidents without delay; the duties apply since August 2025, the Commission gained fining power this August, and this frame matches the headings of the Article 55 record on ai-act-law.eu. The Next Web still caught two snags: Anthropic keeps the alignment review light by claiming no new capability frontier, while the same announcement describes a cyber jump needing frontier-grade safeguards; on price, the company says 10 dollars matches Sonnet 5, yet the outlet had written in June that these were promo prices until August 31 with Sonnet 5 moving to 15 dollars after, so either the hike never came or the comparison uses the promo price. A new Haiku lands within weeks with no date; the week's roundup also counts OpenAI's mid-tier and budget Sol and Luna lines plus the model said to power a coming Meta smart-glasses feature.