Platform · Amérique du Nord

Claude Managed Agents

Claude Managed Agents is Anthropic-managed Claude agent infrastructure for developers and long-running tasks. It lets teams configure, run, and observe agents in persistent sessions with a sandbox, configurable tools, MCP servers, and skills.

ExperimentalAnthropicÉtats-Unis

In a few words

For developers and teams integrating Claude into an application who want Anthropic to operate the agent loop, configured tool execution, and runtime environment.

What you can do with it

Create your own agent

API, CLI, and Console document defining, versioning, and testing an agent.

With development

Automate tasks

Persistent sessions, tools, and SSE events support long-running or asynchronous work, with an integration to build.

With development

Connect an agent to applications or data

MCP servers and custom tools are configurable; your application executes custom tools and returns their results.

With an additional connection

Research the web

web_search and web_fetch are configurable built-in tools; web search is billed when triggered.

With a specific offer

Concrete examples

Define an agent

Create an agent with this system prompt, this MCP server, and these search tools.

Defining an agent with a model, prompt, tools, MCP, and Skills is documented; a session and environment are still required to run it.

Example based on a documented feature

Start a session

In this session, examine this document and use web search if the configuration allows it.

Session events accept documents, text, and images; web tools can be enabled/restricted and their use is billed.

Example based on a documented feature

Access and terms

FR — La facturation est à l’usage: jetons au tarif du modèle, 0,08 USD par heure de session en état running et, si utilisée, recherche Web à 10 USD pour 1 000 recherches. Aucun accès gratuit ni crédit gratuit propre à Managed Agents n’est promis dans les sources examinées. EN — Billing is usage based: model tokens, $0.08 per running session-hour, and, when used, web search at $10 per 1,000 searches. No free access or Managed-Agents-specific free credit is promised in the reviewed sources.

Free plans, trials and paid APIs are separate. Always check the exact offer.

What you can send and get

A session event accepts text, image, or document. API requests are REST with application/json and responses/states can be followed by SSE events; no universal output format is established.

Agents and automation

Agent creation is documented through the API, ant CLI, or Console UI: configure a versioned agent and environment, test a session, then integrate identifiers into your code. This is not a one-click autonomous application.

Go deeper

FR — Une définition d’agent versionnée peut comprendre modèle, prompt système, outils, serveurs MCP, Skills, description, métadonnées et orchestration multi-agent; elle doit être lancée dans une session avec environnement. Les entrées documentées d’un message sont texte, image et document; l’API REST utilise application/json et les événements peuvent être suivis par SSE. Les outils intégrés incluent bash, fichiers, web_search et web_fetch; un outil personnalisé est exécuté par votre intégration, pas directement par Claude. EN — A versioned agent definition can include a model, system prompt, tools, MCP servers, Skills, description, metadata, and multi-agent orchestration; it must be run in a session with an environment. Documented message inputs are text, image, and document; the REST API uses application/json and events can be followed through SSE. Built-in tools include bash, files, web_search, and web_fetch; a custom tool is executed by your integration, not directly by Claude.

textimage inputtool useautomationagent creation

Sources and update

Official Anthropic documentation checked on 13 September 2026: overview, agents, Console, sessions, tools, environments, limits, and pricing. Product is in beta.

https://platform.claude.com/docs/en/managed-agents/overview