AI Agent Autonomously Completes Portal in 24 Hours; iRobot Rebounds with Dual-Robot Design
AI enthusiast CozyBlaze used OpenAI's GPT-6 Astra to autonomously clear Valve's 3D puzzle game Portal by connecting it to SourcePauseTool via the MCP protocol. Operating in a 'pause-analyze-execute' loop, the agent took 24 hours, made 3,336 tool calls, and incurred $571.18 in API fees—the first demonstration of long-horizon autonomous decision-making in a 3D environment without goal drift.
GPT-6 Astra, released on Sept. 3, 2026, delivers the underlying capability: it scores 72.6% on the OSWorld 2.0 benchmark, runs 47% faster than its predecessor GPT-5.6 Sol, and achieves 92.7% accuracy on ScreenSpot-Pro for locating on-screen elements. These metrics mark a generation leap from 'reading screens' to 'operating computers.'
The experiment extends AI reasoning from text and code into simulated 3D space, offering direct validation for spatial reasoning that robotics companies like Figure AI (valued at roughly $39 billion) need. It also accelerates use cases such as digital employees and automated software testing. Analysts see this as a stepping stone toward AI moving from understanding the world to acting on it.
In December 2025, robot vacuum pioneer iRobot filed for Chapter 11 bankruptcy. Its largest contract manufacturer, Shenzhen Shinecon, took full control through a debt-to-equity swap, with iRobot owing Shinecon and Carlyle Group more than $350 million while holding only $24.8 million in cash. The company was delisted from the Nasdaq in February 2026 with a market capitalization below $25 million, marking the fall of a former industry leader to a brand owned by a Chinese manufacturing firm.
At IFA in September 2026, iRobot unveiled the Roomba Duo, a two-body cleaning system designed to work together. The main unit incorporates a 46-newton downforce rolling mop with steam preheating; a secondary unit stands only 83.5 millimeters tall, allowing it to slip under low-profile furniture. The two robots connect via a mechanical deployment arm and share a home map to coordinate cleaning. This architecture departs from the industry's single-all-in-one approach, instead using two specialized machines to overcome the thickness limitations that prevent conventional flagships (typically 95–105 mm high) from reaching under beds and sofas.
Competitive pressure is intense: Chinese brands occupied all top five spots globally in 2025–2026 with a combined share above 70% (Roborock 27%, Dreame 23.7%, etc.), and their marketing budgets reach 3–5 billion RMB annually. Strapped for cash, iRobot cannot outspend rivals on suction power or dock features. The dual-robot strategy is a bet that the competitive battlefield will shift from 'dock comprehensiveness' to cleaning coverage. However, the company faces hurdles: consumer education is costly, the mechanical deployment arm's reliability is unproven, and Chinese rivals can imitate quickly—Dreame's shipments grew 101.9% in 2025.