{"id":68206,"date":"2026-05-19T18:47:13","date_gmt":"2026-05-19T16:47:13","guid":{"rendered":"https:\/\/www.apfelpatient.de\/?p=68206"},"modified":"2026-05-19T18:48:15","modified_gmt":"2026-05-19T16:48:15","slug":"anthropic-claude-managed-agents-mcp-tunnels-karpathy","status":"publish","type":"post","link":"https:\/\/www.apfelpatient.de\/en\/news\/anthropic-claude-managed-agents-mcp-tunnels-karpathy","title":{"rendered":"Anthropic expands Claude Managed Agents and brings in Andrej Karpathy"},"content":{"rendered":"

Anthropic announced a major technical and a personnel development on the same day. Claude Managed Agents received two new enterprise features for privacy and security \u2013 and simultaneously, Andrej Karpathy, one of the most renowned AI researchers, moved from OpenAI to Anthropic.<\/strong><\/p>\n\n\n\n

May 19, 2026, marks a double market day for Anthropic. On the product side, two new features for Claude Managed Agents, MCP Tunnels and Self-Hosted Sandboxes, are being introduced, giving companies greater control over security and data flow. On the personnel side, Andrej Karpathy is announcing his move to Anthropic \u2013 a move of this magnitude rarely seen in the AI industry. Karpathy was a founding member of OpenAI, later served as Director of AI at Tesla, and is now one of the most influential voices in the field. His move comes at a time when Anthropic, with a confirmed valuation of $380 billion following its Series G funding round in February, is among the world's most valuable tech companies \u2013 and is negotiating another round that, according to reports in the New York Times, could raise the company's valuation to as much as $950 billion. The new connectors for creative tools<\/a> and the recent enhancements to Managed Agents demonstrate that Anthropic has clearly accelerated its pace of product updates.<\/p>\n\n\n\n

Two new features for Claude Managed Agents<\/h2>\n\n\n\n

Claude Managed Agents were introduced in April and are designed to significantly simplify the development and operation of AI agents in cloud environments. At the beginning of May, Anthropic expanded the platform with three new features \u2013 including Dreaming, Outcomes, and multi-agent orchestration. Now, the next<\/a> update focuses on security and data privacy.<\/p>\n\n\n\n

MCP Tunnels<\/strong> allow Claude agents to access MCP servers within a private network without exposing them to the public internet. Internal databases, private APIs, knowledge repositories, and ticketing systems become tools that the agent can use\u2014without requiring inbound firewall rules or public endpoints. A lean gateway establishes a single outbound connection, and the traffic is end-to-end encrypted. The feature launches as a Limited Research Preview, so companies must request access before deployment.<\/p>\n\n\n\n

Self-hosted sandboxes<\/strong> offload tool execution to your own infrastructure. Sensitive files, packages, and services remain where your company already has them - either in its own data center or with a managed sandbox provider. The agent loop, responsible for orchestration, context management, and error handling, remains on Anthropic's infrastructure. Support is provided for your own sandbox clients as well as partner solutions from Cloudflare, Daytona, Modal, and Vercel. This feature is currently in public beta and therefore available immediately.<\/p>\n\n\n\n

Why these updates are crucial for businesses<\/h2>\n\n\n\n

These two features address a key obstacle to deploying AI agents in an enterprise context. Anyone granting an agent access to internal systems wants to be able to restrict that access to their own network. Until now, this was only possible to a limited extent with cloud-hosted agents \u2013 now Anthropic delivers a model that takes the security requirements of large organizations seriously.<\/p>\n\n\n\n

Strategically, Anthropic is clearly positioning itself against OpenAI in the enterprise market. While OpenAI has recently expanded its own enterprise paths with Codex and ChatGPT for Intune, Anthropic relies on an architecture that cleanly separates external execution from internal data sovereignty. For many corporations \u2013 especially in Europe \u2013 this is a crucial difference.<\/p>\n\n\n\n

Karpathy is leaving the Eureka Labs solo project and switching to Anthropic<\/h2>\n\n\n\n

Alongside the product announcement comes a major personnel announcement: Andrej Karpathy is joining Anthropic. He himself announced<\/a> his move to X and emphasizes that the next few years will be particularly formative at the forefront of Large Language Models \u2013 he is happy to return to the team and to actively participate in research and development again. He also intends to continue pursuing his passion for AI-powered education.<\/p>\n\n\n\n

Karpathy's career has been unusually fast-paced. He was one of the founding members of OpenAI in 2015 and worked there as a research scientist. In 2017, he moved to Tesla, where he built and led the Autopilot Vision team as AI Director until 2022. He then returned to OpenAI, where he worked on mid-training and synthetic data generation. In 2024, he left OpenAI for a second time and founded Eureka Labs, a startup focused on AI and education. Thus, his move to Anthropic is not the first pivot in his career \u2013 but arguably the most symbolic.<\/p>\n\n\n\n

What the shift reveals about the balance of power in AI<\/h2>\n\n\n\n

The move of an OpenAI founding member to a direct competitor is more than just a personnel announcement. It confirms a shift that has been apparent for months. Anthropic closed a Series G funding round in February 2026 with a post-money valuation of $380 billion \u2013 led by the sovereign wealth fund GIC and the investment firm Coatue. Negotiations are currently underway for a new funding round that, according to reports in the New York Times, could bring the company up to $950 billion. If this valuation were reached, Anthropic would surpass OpenAI for the first time, which was valued at $852 billion in March 2026. At the same time, Anthropic is preparing for a potential IPO \u2013 insiders are citing October 2026 as a target timeframe, though there has been no official commitment from the company so far.<\/p>\n\n\n\n

Karpathy is known for aligning his career moves with substantive questions and research opportunities, rather than conventional career paths. His move to Anthropic now signals two things: the company has clearly established a position in frontier LLM research that is attractive enough for someone like Karpathy, and Anthropic is expanding its research team in parallel with its commercial growth. There's also the historical context: Karpathy isn't the first researcher to leave OpenAI for Anthropic. Dario and Daniela Amodei founded Anthropic in 2021 after leaving OpenAI, and a large part of Anthropic's early team also came from OpenAI.<\/p>\n\n\n\n

A day that shifts the AI landscape<\/h2>\n\n\n\n

Anthropic pulled two levers simultaneously today. The new managed agent capabilities expand its product offerings in an area where companies are particularly focused on data control. The Karpathy acquisition signals that the pendulum is also swinging further in Anthropic's favor in the competition for top AI researchers. Together, these developments paint a picture rarely seen in the industry: a company simultaneously advancing its technology, expanding its workforce, and positioning itself financially to rival or surpass OpenAI \u2013 all on a platform that is becoming significantly more accessible to Apple users with iOS 27. (Image: Shutterstock \/ Stockinq)<\/p>\n\n\n\n