DEV Community

정상록
정상록

Posted on

Claude Managed Agents: Anthropic's New Cloud Agent Platform (Complete Guide)

#agents #ai #claude #cloud

Claude Managed Agents: Anthropic's New Cloud Agent Platform

On April 8, 2026, Anthropic released Claude Managed Agents in public beta — a composable API set for building and deploying cloud-hosted AI agents. The promise: go from prototype to production in days, not months.

The Problem It Solves

Building production agents is mostly infrastructure work:

  • Security sandboxing
  • Authentication & authorization
  • Tool execution management
  • State persistence across failures
  • Error recovery & retry logic
  • Session management
  • Monitoring & logging

Managed Agents handles all of this behind the API. You focus on agent logic.

Architecture: 4 Core Concepts

Concept What It Is
Agent Model + system prompt + tools + MCP servers
Environment Cloud container (packages, network, mounts)
Session Agent instance execution unit (hours-long)
Events SSE streaming for app-agent communication

Quickstart (5 minutes) 

import anthropic

client = anthropic.Anthropic()

# Define agent
agent = client.agents.create(
    model="claude-sonnet-4-6",
    name="Coding Assistant",
    system="You are a helpful coding assistant.",
    tools=[{"type": "agent_toolset_20260401"}]
)

# Create environment
environment = client.environments.create(
    name="Development",
    packages={"python": "3.12", "nodejs": "22"}
)

# Start session
session = client.sessions.create(
    agent_id=agent.id,
    environment_id=environment.id
)

# Send task and stream results
client.sessions.send_event(
    session_id=session.id,
    event={"type": "user", "content": "Write fibonacci.py"}
)

for event in client.sessions.stream_events(session_id=session.id):
    print(event)
Enter fullscreen mode Exit fullscreen mode

Beta header managed-agents-2026-04-01 is set automatically by the SDK. The ant CLI was also released for command-line interaction.

Key Features

Long-Running Sessions

Sessions run autonomously for hours. Survives network failures and client disconnects. Progress and results are preserved.

Multi-Agent Orchestration (Research Preview)

Agents can spawn and coordinate other agents for parallel processing. Currently requires separate access request.

Built-In Governance

Scoped permissions, identity management, and execution tracking are all native. Every action is auditable.

Messages API vs Managed Agents

Messages API Managed Agents
What Direct model prompting Managed agent harness
Best for Custom loops, fine control Long-running tasks, async work
Infrastructure You build it Anthropic manages it

Pricing

  • Token cost: Standard Claude API pricing
  • Session runtime: $0.08/session-hour (active runtime only)
  • Rate limits: 60 creates/min, 600 reads/min

Early Adopters

Companies already in production:

  • Notion — workspace delegation with Custom Agents
  • Rakuten — department-specific agents deployed in one week
  • Sentry — bug detection to PR creation, fully automated
  • Atlassian — Jira task triggers agent execution

Performance

Internal testing: up to 10-point improvement in structured task success rate vs standard prompting. The harness itself enhances agent performance, especially on harder problems.

What This Means for Developers

The shift is clear: stop building infrastructure, start building agent logic. Security, auth, state management, long-running execution — it's all behind the API now.

The question is no longer "how to build agents" but "what to build agents for."

Resources:

Top comments (0)