Topic
Agents that see, click and type in graphical applications and operating systems rather than only editing text.
Pine AI's cloud computer for agents resolved 27.4% of SaaS-Bench tasks in Pine's own test, below two rivals at 31.1% and 29.2%. Its higher step score and much lower token bill come from whole systems running different models, so the computer's own contribution has not been measured.
Reality
Apple pulled a Mac mini and Mac Studio refresh forward to August 25 to meet buyers that, according to The Information, no engineering, developer relations or enterprise AI function inside the company was assigned to serve.
Reality
OpenAI and Ironclad published a case study on turning production contract workflows into training data and evals for computer-use agents. Scores from such an eval carry over to other software only as far as it shares contract approval's roles and branching.
Reality
Brett Adcock says Hark launches this week and gives its first 100,000 sign-ups the paid plan free. What each user connects at sign-up decides how far one agent reaches into their inbox, calendar and payment accounts.
Perspective Coverage
3 publishersReality
H Company's open-weight Holo4 27B handles GUIs, code, MCP servers and REST APIs in one model, scoring 61.7% on OSWorld 2.0 at an estimated $1.22 per task. The choice between specialists happens during training, so agent teams could drop the router from their stack if the scores hold on their own workloads.
Reality
OpenAI is selling Dots, a cartoon agent that runs apps on a virtual machine, first to top-tier accounts like the $100-a-month Pro plan. Its errands for The Verge kept ending in human takeovers, so desk work in a team's own apps is the case to test.
Reality
Manus 2.0 lets agents fire on email and Slack events and stay online on dedicated cloud computers, with a harness Manus says cost 32% less in one test. Manus has not described that test, so the 32% is a figure for its own workload until someone reruns it elsewhere.
Reality
Instinct raised $1 billion at a $10 billion valuation, four times the $2.5 billion it was valued at a month earlier. Its agent runs errands by operating other companies' websites, apps and phone lines, so single-task apps end up serving software that acts for the user.
Perspective Coverage
3 publishersReality
beans-picker cut a median macOS agent task from $0.388 to $0.060 in its author's benchmark by replacing full UI-tree reads with short candidate lists. Whether that carries over depends on how much of your own agent's bill goes to reading the window.
Reality
The always-on agent product opened in beta on August 11. The docs cap accounts at 50 bots and chats, give every bot on an account the same computer, and skip Linux desktop entirely.
Reality
Generating the interface as video instead of rendering it from code is a serious research bet. Runway's own preview still lists legible text and long-session coherence as open problems, which is roughly where ordinary interfaces begin.
Perspective Coverage
5 publishersReality
Sensor Tower counted more than 900,000 downloads of Meta's Muse in its first week. Wired's reviewer found an agent that browses competently while its ideas tab keeps pitching a live link to the bank.
Reality
Almost every announcement at the hour-long keynote tied back to Meta's personal agent, including glasses from $449 and a Mac that Muse can operate while the user is away. The metaverse barely came up.
Reality
The company known for a pocket gadget now ships a cloud agent that installs on up to five Windows, macOS or Linux machines and runs on whatever model subscription the user already pays for. Four rival agents got there first.
Reality
Sol lists at $2 and $10 per million tokens and Luna at $0.10 and $0.50. The per-task savings OpenAI published come mostly from the lower price, and the cheaper tier scores below its predecessor on computer use.
Reality
Cua's pitch is an OS-level driver that delivers clicks and keystrokes in the background on macOS, Windows and Linux. The post hedges it with "where supported by the platform". A team has to test that clause before designing around it.
Reality
simframe's author rebuilt iOS Simulator perception as an always-on daemon after watching Claude Code take a fresh screenshot before every tap. The study he cites puts 75-94% of task time inside model calls.
Reality
Cua published a 706,048-parameter model, its training data and its driver integration under MIT, so the claim that a small specialist can take over the per-field decision step is testable by anyone with forms.
Reality
OpenAI's GPT-6 Astra video shows the model driving Blender on voice commands. The 3D and game-development communities spent the next two days telling OpenAI to stay away from human-made art, and a Fast Company column argues that better demos now land worse.
Reality
Anthropic's support note says the mode picker goes away in a staged rollout, and accounts that move over cannot return to separate Chat and Cowork options. Connected apps stay live while a task runs.
Reality
Invest · September 12, 2026 · Confirmed2 publishers
Product · September 3, 2026 · One report1 publisher