The AI giant Anthropic is expanding its autonomous agent functionalities to mobile devices and the web. Claude Cowork now operates in the background as well.
Starting now, Max subscribers can use the beta version of Claude Cowork on smartphones and in browsers. The standout feature: AI tasks synchronize across all devices and continue even when the device goes offline. Moreover, until August 5th, Anthropic is doubling usage limits for these functionalities.
What Users Really Do with AI Agents
The rollout is based on an analysis of 1.2 million sessions between May 11 and 31 of this year. The statistics clearly show that business processes dominate usage at 33.4%, followed by content creation at 16.4% and software development at 8.7%.
With the new mobile accessibility, users can plan tasks and manually check the results generated by AI agents. This shift away from solely desktop access enables more enduring, autonomous workflows, allowing agents to operate independent of the user’s active screen time.
As the EU AI Act sets the legal framework for implementing such AI systems in businesses starting August 2024, companies must keep themselves informed about requirements and responsibilities. A comprehensive implementation guide is available to provide an overview of all relevant information for IT departments.
Nvidia’s New Chip Generation Fuels Development
In parallel to software advancements, hardware is delivering the necessary computational power. On July 7, Nvidia unveiled the Vera CPU, equipped with 88 Olympus cores. This processor is tailored for AI factories, offering 1.8 times higher performance per core with 40% lower latency compared to x86 architectures.
Energy efficiency is equally impressive: it boasts three times more memory bandwidth at half the power consumption. This is achieved through a monolithic die with up to 1.2 TB/s bandwidth via LPDDR5x.
Software developments are also in the works. Couchbase recently announced the general availability of its AI Data Plane, a system that unifies agent storage, catalogs, and MCP servers, supporting frameworks such as LangGraph, CrewAI, and LlamaIndex. Researchers from Harvard SEAS introduced the Orla framework at CAIS 2026, optimizing multi-model AI workflows based on cost, speed, and accuracy.
Security Vulnerabilities Threaten Corporate Data
However, these new capabilities also bring risks. On July 7, Noma Security uncovered a critical vulnerability in GitHub’s Agentic Workflows. The flaw, dubbed GitLost, allows attackers to manipulate an AI agent through a public GitHub issue, potentially extracting sensitive data from private repositories.
A proof of concept demonstrated how the agent posted a private README file as a public comment, alarming given that no programming skills were required for the exploit. At the time of disclosure, there were no fixes or documentation updates from GitHub.
While new AI technologies may boost efficiency, they simultaneously introduce new cyber risks and legal obligations for entrepreneurs. A free Cyber Security Report is available to help identify and mitigate these vulnerabilities proactively.
Industry Emphasizes Standards and Security
The industry is responding with new integration solutions. Salesforce and AWS announced a partnership on July 6th that connects Agentforce with Amazon Agentcore. This integration utilizes the open MCP standard, allowing AI agents to navigate between different cloud environments and interact with backend systems via OAuth2.
On the governance front, HPE is collaborating with Trustwise to integrate real-time policies into private cloud AI environments. Concurrently, new standards have come into effect – China’s first trust standard for AI agents has been in place since June 11, 2026. An evaluation by AgentRisk mapped 2.3 million agents, revealing that 99.19% were deemed healthy, while only 0.16% were inactive.
The trend toward experimenting with agents continues unabated. A McKinsey survey from 2025 found that 62% of respondents had already experimented with AI agents, even before the current wave of platform expansions began.
