AI startup Anthropic has introduced a new environment for deploying complex and long-running agent tasks called Claude Managed 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.”

Read more: https://t.co/YYaEub2QGV

— Anthropic (@AnthropicAI) April 8, 2026

This solution is a hosted service on the Claude platform that enables the management of long-term agents through a set of basic interfaces.

The system can sequentially execute tasks, utilize tools, run code, edit files, interact with external services, and continue functioning even after failures.

The Problem

Previously, developers faced challenges when launching AI agents:

  • Digital assistants would forget context, necessitating resets;
  • The model would prematurely terminate tasks, requiring special "workarounds";
  • The neural network struggled with long-duration assignments.

For instance, Claude Sonnet 4.5 could finish a task prematurely as it approached the context window limit.

The Solution

The cloud tool Claude Managed Agents handles all background processes, including memory management and error recovery.

The main architectural concept of this new solution is the separation of the agent into independent components. Previously, everything operated in a single environment:

  • The model itself and its invocation logic;
  • Code execution;
  • Memory and session states;
  • Access and tokens.

Anthropic has divided the system into three main blocks:

  1. Session — a log of all actions and events. It serves as the agent's long-term memory, recording the history of steps and allowing for bypassing context length limitations.
  2. Control Logic — the "conductor" that invokes Claude, determines which context to present to the model, and manages the operational cycle.
  3. Execution Environment — where the LLM can run code, edit files, and interact with its surroundings.

This structure allows each part to function autonomously—local failures in one process no longer lead to the closure of the entire session.

Anthropic positions Claude Managed Agents as a resilient layer of infrastructure that will endure the evolution of models. It acts as an operating system for AI agents.

Previously, the company developed a new LLM called Claude Mythos, but withdrew it from public release due to high security risks.