← Back to feed

Anthropic launched Claude Managed Agents into public beta on April 8, 2026, a suite of composable APIs that handles the infrastructure layer of AI agent deployment — sandboxing, state management, tool execution, credential handling, error recovery — so developers do not have to build it themselves.

The pitch is compression: what typically takes months of DevOps work can now be done in days. That is a real problem. Building a production-grade AI agent means constructing secure sandboxes, wiring up tool orchestration, managing context across long sessions, and ensuring reliability when containers die mid-run. Most teams building agents are spending more time on scaffolding than on the actual agent logic.

What It Does

Managed Agents introduces three components: the “brain” (Claude plus an orchestration harness that decides when to call tools and how to manage context), the “hands” (sandboxed execution environments for tools including Bash, file operations, web search, and MCP servers), and the “session” (a durable event log that persists through disconnections so long-running tasks survive infrastructure failures).

Sessions can run autonomously for hours. Outputs persist even when the connection drops. If a container dies, the harness treats it as a tool-call failure and recovers. The architecture is designed for the use cases where the Messages API breaks down: automated code review across large repositories, multi-step research synthesis, backend data-processing pipelines.

Multi-agent coordination — where one Claude agent spawns and directs up to 10 sub-agents in parallel — is available in research preview. Anthropic reports up to 10 percentage points of task-success improvement in structured file generation tests when using multi-agent spawning versus single-agent prompting.

Access is enabled by default for all Claude API accounts. Developers add a single beta header (managed-agents-2026-04-01) to requests; the SDK sets it automatically. Rate limits at launch: 60 create operations per minute, 600 read operations per minute per organization.

Pricing

$0.08 per session-hour for active runtime, plus standard Claude API token costs. For long-horizon tasks that would otherwise require a dedicated server running 24/7, this pay-as-you-go structure significantly undercuts most in-house infrastructure costs. The economics favor bursty, episodic workloads — a code review job that runs for 90 minutes every night is cheap; a continuously running agent is less so.

Early Adopters

The production roster at launch is credible and diverse:

• Notion — Custom Agents let teams delegate open-ended tasks (coding, slides, spreadsheets) without leaving the workspace
• Rakuten — Deployed specialist agents across product, sales, marketing, and finance within a week per deployment, with Slack and Teams integration
• Sentry — Connected its Seer debugging tool to a Claude agent that writes patches and opens pull requests; flagged bug to reviewable fix in a single flow
• Asana — Built AI Teammates that operate alongside human teams on projects
• Atlassian — Building agents that let developers assign tasks directly from Jira

The breadth across productivity, project management, e-commerce, and developer tooling suggests the deployment friction problem is real and widespread across enterprise software categories.

What This Means

Anthropic is shifting from model vendor to infrastructure platform. The company is now competing with AWS and Azure for the agent runtime layer — not just the inference layer. That is a significant expansion of scope, and it lands at the same moment that Claude Cowork reached general availability (April 9), creating a two-pronged strategy: consumer workspace tool for end users, developer infrastructure for builders.

The risk that enterprises will flag is obvious: routing all agent runtime through Anthropic’s servers means the “brain” is always on Anthropic’s infrastructure, even when VPC connectivity is available for the “hands.” For sensitive workloads, that raises governance questions that Managed Agents does not fully resolve.

For everyone else, the calculus is straightforward: if you are manually splitting long-horizon tasks across multiple Claude sessions today, this beta is worth evaluating immediately.

Key Numbers

• Launch date: April 8, 2026 (public beta)
• Pricing: $0.08/session-hour + standard token costs
• Rate limits: 60 create/min, 600 read/min per org
• Multi-agent spawning: up to 10 sub-agents (research preview)
• Task-success improvement: up to +10 points vs. single-agent prompting
• SDK support: Python, TypeScript, Java, Go, Ruby, PHP
• Early production customers: Notion, Asana, Sentry, Rakuten, Atlassian