Anthropic’s Agent Push
Anthropic has opened Claude Managed Agents in public beta, making it easier for teams to deploy long-running, orchestrated AI agents for production use (x.com). At the same time the company is running Project Glasswing — a defensive effort to detect software vulnerabilities using its models — underscoring that Anthropic is pushing both agent capabilities and safeguards in parallel (x.com).
Most artificial intelligence agents still break the moment they have to do real work for more than a few minutes. Anthropic’s new Managed Agents service is a bet that companies will pay to stop stitching together their own agent plumbing. (platform.claude.com) A managed agent is basically rented infrastructure for a long-running software worker. In Anthropic’s setup, a company defines the model, tools, and rules, and Anthropic runs the session, the container, and the event stream. (platform.claude.com) The product is in public beta now, and the application programming interface uses a beta header dated 2026-04-01. The quickstart shows three separate pieces: an agent, an environment, and a session. (platform.claude.com) That structure solves a boring problem that has slowed down most agent projects. The “agent” is the brain, the “environment” is the workshop, and the “session” is the actual job in progress. (platform.claude.com) Anthropic is also shipping this with prebuilt tools instead of forcing developers to wire up every screwdriver by hand. Its agent toolset includes bash commands, file operations, web search, and other built-in actions. (platform.claude.com) This did not come out of nowhere. Anthropic spent 2025 and 2026 publishing research on long-running agents, multi-agent orchestration, and “harnesses,” which is its term for the control layer that keeps an agent on task over hours or days. (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) At almost the same moment, Anthropic launched Project Glasswing with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. That project is aimed at securing critical software before attackers use similar models first. (anthropic.com) Anthropic says its unreleased Claude Mythos 2 Preview model has already found thousands of high-severity vulnerabilities, including flaws in every major operating system and web browser. That is why Glasswing is framed as a defensive program, not a normal product launch. (anthropic.com) The company is putting money behind that claim. Anthropic says it will provide up to $100 million in usage credits for Glasswing work and another $4 million in direct donations to open-source security groups. (anthropic.com) Put together, these two launches show Anthropic trying to own both halves of the agent era. It wants to sell the machinery that lets companies run autonomous software workers, while also arguing that the same class of models is now powerful enough to force a full-scale defensive push in cybersecurity. (platform.claude.com) (anthropic.com)
