LLM Agents Usher In Agentic Web as GPT-6 Astra Hits 73% on OSWorld 2.0
Updated
Updated · MIT Technology Review · Nov 2
LLM Agents Usher In Agentic Web as GPT-6 Astra Hits 73% on OSWorld 2.0
3 articles · Updated · MIT Technology Review · Nov 2
Summary
Autonomous software agents are moving from concept to mainstream use, with tools such as Meta’s Muse, OpenAI’s Dots and Anthropic’s Claude Duet now handling tasks from shopping to inbox triage.
Large language models made that possible by reading ordinary web pages without special semantic tags, while newer multimodal systems can interpret screenshots, click menus and complete multistep software tasks.
GPT-6 Astra’s 73% score on the OSWorld 2.0 benchmark suggests agents are nearing human-level web navigation, reducing the need for custom agent-only infrastructure and making the existing web accessible to them sooner than expected.
That shift is already reshaping online business models: Time is publishing agent-friendly articles and selling ads to crawlers, while publishers face traffic and revenue pressure as AI summaries cut click-throughs by about 40%.
The longer-term outcome remains unsettled, with connectors and closed platforms offering more safety and control, but raising the risk that a few dominant companies—or even agents themselves—could define how the web evolves.
Will AI agents free us from tedious tasks, or trap us in invisible digital walled gardens controlled by tech giants?
If AI agents do all our shopping and reading, how will websites survive when human eyeballs disappear?
When a malicious website tricks your AI assistant into draining your bank account, who is legally responsible?
Surpassing Human Baselines: GPT-6 Astra, the Agentic Web, and the Economic Shockwave of 2026
Overview
In September 2026, OpenAI launched GPT-6 Astra, a powerful AI model trained on over 100,000 GPUs, marking a major leap in autonomous computer use and enterprise workflows. Astra’s advanced capabilities led to high pricing and forced companies to rethink ROI and management structures, as AI agents began replacing routine tasks and coordinator roles. However, Astra’s autonomy also brought critical cybersecurity risks, prompting OpenAI to restrict its use and withhold certain features. As organizations adopted these agents, new regulatory and safety challenges emerged, including unreliable evaluation methods and increased data breaches, ultimately leading OpenAI to pause further releases due to unresolved security concerns.