Anthropic ships managed agent harness

Anthropic has moved its agent product from recipes and engineering blog posts into a managed platform service. The company’s documentation and release notes show Claude Managed Agents in public beta, with memory added on April 23, webhooks and multiagent sessions added on May 6, and access through Claude Platform on AWS added on May 11. Anthropic’s own engineering write-up describes Managed Agents as a hosted service for “long-horizon agent work” built around stable interfaces for sessions, harnesses and sandboxes. In that telling, the product is not just a model endpoint. It is a managed execution layer for asynchronous work that can keep running while Anthropic changes the harness underneath. ### What exactly did Anthropic ship? Anthropic’s release notes say Claude Managed Agents are a “fully managed agent harness” with secure sandboxing, built-in tools and server-sent event streaming. (platform.claude.com) The service sits inside the Claude Platform rather than as a standalone open-source framework. The April 8 engineering post says the company built the service because harnesses “encode assumptions” that can become obsolete as models improve. (anthropic.com) Anthropic said it wanted interfaces that “outlast any particular implementation,” including the ones it runs today. ### Why is the harness itself part of the product? Anthropic’s engineering team said prior harness designs had to compensate for model-specific behavior. (platform.claude.com) One example in the post described Claude Sonnet 4.5 ending tasks early as its context window filled up, leading Anthropic to add context resets; the same workaround became unnecessary on Claude Opus 4.5. That account matters because Anthropic is selling more than inference. (anthropic.com) The company said Managed Agents virtualize three components — session, harness and sandbox — so implementations can change without breaking the application layer above them. ### Which features make this a durable, long-running service? Anthropic added memory for Claude Managed Agents in public beta on April 23, according to the platform release notes. (anthropic.com) On May 6, the company added multiagent sessions, outcomes, webhook support, background credential refresh for vaults, and more session and event filtering. Those features address the mechanics of long-running work. Memory lets an agent retain context across work, webhooks let outside systems subscribe to session and vault lifecycle events, and filtering tools help operators inspect sessions by status or event type, according to the same release notes. (anthropic.com) ### What do the deprecation notices tell customers? Anthropic’s release notes show model lifecycle management is part of using the platform. (platform.claude.com) On April 30, the company retired the 1 million-token beta header for Claude Sonnet 4.5 and Claude Sonnet 4, and said requests above the standard 200,000-token window would error on those models. Anthropic told customers to migrate to Claude Sonnet 4.6 or Claude Opus 4.6 if they wanted a 1 million-token context window without the beta header. (platform.claude.com) That is a concrete example of the maintenance burden for teams building on managed agents: the harness may be abstracted, but model upgrades and retirements still require planning. ### How does AWS fit into the rollout? Anthropic said on May 11 that Claude Platform on AWS gives customers access to the Messages API, Files API, Message Batches API, Claude Managed Agents, Agent Skills, code execution and tool use through AWS endpoints. (platform.claude.com) AWS handles billing and IAM-based access control, while Anthropic operates the inference stack. The AWS documentation also says data may not reside in AWS and inference may route to Anthropic’s primary cloud, while Zero Data Retention is available on request. (platform.claude.com) For enterprise buyers, that makes Managed Agents part of a broader infrastructure and procurement conversation, not just an API feature list. That framing is an inference from Anthropic’s architecture and billing documents. ### Where is Anthropic taking the product next? Anthropic’s May 5 announcement for financial-services agents said ten ready-to-run templates would ship as plugins in Claude Cowork and Claude Code and as cookbooks for Claude Managed Agents. The company positioned those templates for tasks such as pitchbooks, KYC screening and month-end close work. Anthropic’s public-beta path is visible in its own documents: memory entered beta on April 23, multiagent sessions and outcomes on May 6, and AWS access on May 11. (platform.claude.com) The next concrete milestones are likely to appear first in the Claude Platform release notes and the Managed Agents documentation, where Anthropic is already posting feature launches and migration requirements. (platform.claude.com) (anthropic.com)