Claude Agents Public Beta

Anthropic moved Claude Managed Agents into public beta on April 8, 2026, and the pitch is simple: companies can now tell Claude what job to do, add tools and guardrails, and let Anthropic run the agent in the cloud instead of building the plumbing themselves. (claude.com) That plumbing is the part most demos skip. Anthropic says production agents need sandboxed code execution, checkpointing, credential management, scoped permissions, and tracing before a customer ever sees a useful answer. (claude.com) Claude Managed Agents is Anthropic’s answer to that bottleneck. The service gives developers a built-in orchestration harness that decides when to call tools, how to manage context, and how to recover from errors while the developer defines the task, tools, and rules. (claude.com) Anthropic is not just selling a chatbot here. Its own documentation says Managed Agents is for jobs that run for minutes or hours, keep state across interactions, and need secure containers, network access, and persistent file systems. (platform.claude.com) The tool bundle is closer to a remote worker than a text box. Managed Agents can run shell commands, read and edit files, search the web, fetch web pages, and connect to outside services through the Model Context Protocol, which Anthropic introduced in November 2024 as a standard for linking artificial intelligence systems to business software and data. (platform.claude.com) (anthropic.com) That last piece is why names like Notion and Asana keep coming up around this launch. The whole point of the Model Context Protocol is to let an agent reach into the places where work already lives, like project trackers, document stores, and internal tools, instead of forcing employees to copy everything into one prompt. (anthropic.com) Anthropic has been walking toward this for more than a year. In November 2025 it added advanced tool-use features so Claude could discover tools on demand and avoid stuffing tens of thousands of tokens of tool definitions into every request, which is one of the main reasons big agent systems get slow and expensive. (anthropic.com) It also spent 2025 pushing deeper into enterprise controls. In August 2025 Anthropic added new business-plan admin features and a Compliance Application Programming Interface so companies could audit usage and manage access, which are the kinds of controls procurement teams ask about before approving software that can touch internal systems. (anthropic.com) The pricing shows who Anthropic expects to buy this first. The company charges standard Claude token rates plus $0.08 per active session-hour, and web search adds $10 per 1,000 searches, which makes this look less like a consumer feature and more like infrastructure for teams that already spend real money on software operations. (thenewstack.io) Some of the flashier pieces are not fully open yet. The New Stack reports that advanced memory tooling, multi-agent orchestration, and self-evaluation features remain in limited research preview, so the public beta is broad but not the whole roadmap. (thenewstack.io) What Anthropic is really launching is a hosted middle layer between a model and a company’s software stack. If that layer works, businesses get agents without building the runtime themselves; if it fails, Anthropic becomes responsible for one of the hardest parts of the agent business: keeping autonomous software useful, cheap, and under control. (claude.com) (platform.claude.com)