Anthropic Launches Claude Managed Agents — A Platform for Running Long-Lived AI Agents

Anthropic Unveils Hosted Environment for Complex Agent Tasks

AI company Anthropic has launched Claude Managed Agents, a cloud-based service built for executing complex, long-running agent workloads. The platform operates on top of Claude's infrastructure and gives developers a set of core interfaces to manage persistent agents.

"New on the Engineering Blog: Building Managed Agents—our hosted service for long-running agents—meant solving an old problem in computing: how to design a system for 'programs as yet unthought of.'" — Anthropic (@AnthropicAI), original post

The new tool can sequentially process instructions, leverage external tools, execute code, edit files, interact with third-party services, and continue functioning even after encountering failures.

Why This Matters

AI agents represent one of the industry's most active frontiers, but deploying them for sustained, long-running tasks has been plagued by technical limitations. Developers consistently ran into three core problems:

• Digital assistants would lose context, forcing a full session reset;
• Models would terminate tasks prematurely, requiring workarounds and hacks;
• Neural networks generally struggled with extended assignments.

Specifically, Claude Sonnet 4.5 was known to prematurely end tasks as it approached the context window limit. Claude Managed Agents aims to eliminate these barriers, paving the way for more reliable autonomous agent operations.

A Three-Component Modular Architecture

The central architectural principle behind the platform is decomposing the agent into independent components. Previously, everything lived in a single monolith: the language model itself, call logic, code execution, session state, and access tokens.

Anthropic split the system into three standalone blocks:

• Session — a log of all actions and events. This component serves as long-term memory, recording the history of steps and enabling the agent to work around context length limitations.
• Orchestration logic — the "conductor" that calls Claude, determines what context to feed the model, and manages the execution cycle.
• Execution environment — the sandbox where the LLM actually runs code, modifies files, and interacts with external systems.

This design ensures each element operates independently — a local failure in one process no longer crashes the entire session. The cloud service handles background operations including memory management and error recovery.

An Operating System for AI Agents

Anthropic positions Claude Managed Agents as a durable infrastructure layer designed to outlast the evolution of individual models. In essence, the company is building an operating system for AI agents — one that abstracts away the internal complexity and instability of individual components.

Earlier, Anthropic developed a new LLM called Claude Mythos but chose not to release it publicly due to significant safety concerns.