What happened
Simular announced Box on November 12, 2026, describing it as a cloud computer built for agents performing computer tasks. The company frames the product around a familiar problem: agents need access to files, webpages, applications and terminals, but setting up a custom sandbox and computer-use stack can take substantial effort. Simular also says that users assigning agents demanding desktop work found that those tasks consumed usage faster than expected.
Box combines shell, browser and desktop control, all driven by Simular’s computer-use runtime. The company says that logins, files and applications live on Box and are shared among agents. Its example is that Claude can start a job and ChatGPT can finish it. Users can also upload and download files between Box and their agents. This makes the product less like an isolated session for one assistant and more like a common computer that multiple agents can use.
Simular says Box can be used with chat assistants including Claude, ChatGPT and Gemini, and names OpenClaw, Hermes and Claude Code as agents that can use an always-ready cloud computer. For coding work, the company says a desktop can be shared across Claude Code, Codex and Cursor without setup. The announcement directs interested users to box.simular.ai. It does not state a price, usage quota, supported operating system, or detailed setup process.
The announcement makes a performance claim: Simular says its agent reaches state-of-the-art performance on the OSWorld benchmark with “best cost-efficiency,” and describes Box as “10x faster and cheaper.” It gives no comparison baseline, benchmark scores, workload definition, or cost calculation alongside those claims. The company says beta users replaced the sandbox used by some personal agents, while others turned a free agent account into a personal agent; it does not identify those users or quantify the results.
Simular presents Box as a way for agents to do work on a computer while people retain a cooperative role, rather than as a replacement for people. The product announcement is a launch description, not an independent evaluation of reliability, security or performance.
What the archive adds
New Runtime’s earlier coverage described Manus 2.0 as combining agent execution, creative work and persistent hosting in one platform (September 29, 2026 note). Another note covered an isolated Windows desktop for GUI automation (November 8, 2026 note). Box is a new product announcement in the broader thread of computer environments for agents, with its stated emphasis on one shared computer across agents.
Trend signal
Box materially supports agent-ready software exposes capabilities: Simular describes a computer environment through capabilities agents can use—shell, browser and desktop control, plus shared files and applications. The announcement does not establish how those capabilities are permissioned or exposed in practice, so it supports the product-boundary idea more than any claim about safe composition.
Why now
Simular says it built Box after hearing from users who found custom sandbox setup too demanding and desktop-heavy agent tasks unexpectedly costly in usage. That explanation connects the launch to a practical adoption problem, though the company does not give dates or measurements for the user feedback.
Why it matters
A shared computer changes the operational question from “which agent owns this session?” to “what work and state can multiple agents safely share?” If the product works as described, teams could avoid recreating application access and files for each assistant, and hand a task between agents without exporting every intermediate artifact manually. That could be useful where a chat agent handles one step and a coding agent or desktop agent handles the next.
The shared context is also a boundary that teams will need to understand, not merely a convenience. Persistent logins and files can make agent work smoother, but they raise practical questions about permissions, separation between jobs, auditability and cleanup. Teams evaluating Box should therefore test it with non-sensitive workflows first, checking which agents can reach which files and accounts and how state persists. Simular’s claims about speed, cost and benchmark performance are promising descriptions, not enough on their own to justify a production migration.
Limits and uncertainty
Simular does not publish benchmark scores, test conditions, a named comparison baseline or the calculation behind “10x faster and cheaper” in the announcement. Its beta examples are not quantified, and the participating users are unnamed. The announcement also does not specify price, quotas, availability limits, security controls, data retention, isolation guarantees or service reliability. Those details matter for teams deciding whether a shared environment is suitable for real work. No independent performance or security evaluation is established here.
Reader takeaway
Box’s central proposition is a persistent computer that multiple agents can share, rather than a separate sandbox for each assistant. That may reduce setup friction and make handoffs easier, but builders should treat the performance claims as unverified and evaluate access controls, persistence and cost on a low-risk workflow before relying on it.