When would you use Claude Managed Agents?

Customer Support Ticket Resolution: A support team routes incoming tickets to a Managed Agent that retrieves customer account data via an MCP-connected CRM, drafts a resolution, and escalates to a human only when confidence is low. Session state preserves the full ticket thread across multiple turns without re-sending context each time.

When would you use Claude Managed Agents?

Month-End Financial Reconciliation: A finance team triggers an overnight agent session that reads transaction exports from cloud storage, reconciles ledger entries against bank statements, flags discrepancies, and writes a structured summary report—all within a single stateful session that can run for hours without developer supervision.

When would you use Claude Managed Agents?

Competitive Pricing Research: A product team defines an Outcome rubric specifying the desired pricing matrix format. A lead agent delegates vendor-specific research sub-tasks to up to 20 specialist subagents running in parallel, then a grader loop evaluates the synthesized output against the rubric and requests revisions until the artifact meets the defined criteria.

← ContentsClaude API · advanced

Claude Managed Agents

Claude Managed Agents is a hosted, server-side infrastructure service from Anthropic that runs autonomous Claude agents on your behalf. Instead of developers building and maintaining their own agent harnesses—provisioning containers, managing state loops, and securing code execution environments—Anthropic provides a turnkey cloud ecosystem. Within this managed environment, Claude can read and write files, execute shell commands, conduct web searches, maintain persistent conversational history across multi-hour or multi-day tasks, and coordinate with sibling agents in multiagent workflows. The service exposes three primary abstractions: Agents (reusable, versioned configurations that define a model, system prompt, and permitted toolset), Environments (secure execution sandboxes, either Anthropic-managed cloud containers or self-hosted infrastructure), and Sessions (stateful runtime instances that bridge an Agent to an Environment and persist an event log). All three resources are created and managed through REST API endpoints, and the session execution loop is observed via Server-Sent Events (SSE) streaming. As of early 2026, Claude Managed Agents is in public beta and available through the direct Claude API and Claude Platform on AWS. All endpoints require the 'managed-agents-2026-04-01' beta header. Billing occurs on two dimensions: tokens consumed (at standard model rates, with prompt-caching multipliers applied) and session runtime, which replaces the separate container-hour billing used in earlier execution models.

When you’d use it

◆Customer Support Ticket Resolution — A support team routes incoming tickets to a Managed Agent that retrieves customer account data via an MCP-connected CRM, drafts a resolution, and escalates to a human only when confidence is low. Session state preserves the full ticket thread across multiple turns without re-sending context each time.
◆Month-End Financial Reconciliation — A finance team triggers an overnight agent session that reads transaction exports from cloud storage, reconciles ledger entries against bank statements, flags discrepancies, and writes a structured summary report—all within a single stateful session that can run for hours without developer supervision.
◆Competitive Pricing Research — A product team defines an Outcome rubric specifying the desired pricing matrix format. A lead agent delegates vendor-specific research sub-tasks to up to 20 specialist subagents running in parallel, then a grader loop evaluates the synthesized output against the rubric and requests revisions until the artifact meets the defined criteria.
◆Secure Enterprise Code Execution with Private MCP — An enterprise security team needs Claude to audit internal codebases that cannot leave the corporate network. Using self-hosted sandboxes and MCP tunnels (Research Preview), agent execution runs entirely inside the corporate boundary while credentials are injected from Vaults—never exposed to the model's context window.
◆Persistent Multi-Day Coding Workspace — A developer opens a long-lived agent session configured as a cloud coding assistant. Unlike stateless API calls, the container retains file edits, installed packages, test results, and conversation history across multiple days, functioning as a durable remote development environment.

What changed recently

◆2026-04 — Memory for Claude Managed Agents entered public beta under the managed-agents-2026-04-01 header. Agents can now persist learned facts and summaries across sessions without polluting original session transcripts.
◆2026-04 — Multiagent sessions and Outcomes entered public beta. A lead orchestrator agent can now spawn up to 20 specialist subagents in parallel and use an independent grader loop to enforce output rubrics automatically.
◆2026-04 — Self-hosted sandboxes became available as an alternative to Anthropic-managed cloud containers, enabling organizations with data-residency or compliance requirements to run tool execution on their own infrastructure while keeping state management in the Claude Platform.
◆2026-04 — MCP tunnels entered Research Preview, allowing agents to connect to MCP servers running inside private corporate networks without exposing those servers to the public internet.

This is the short version

The full chapter has three worked examples, the common pitfalls, and the workflow that makes it pay — plus the other 84 features, kept current.

Get Claude Master — $97 →