Anthropic just removed the hardest part of deploying AI agents: the infrastructure. Here's what SMBs need to know about Claude Managed Agents, pricing, and why this matters for your AI roadmap.
Article text
Building an AI agent is easy. Deploying it to production is where most projects die.
That's the gap Anthropic just filled with Claude Managed Agents, a public beta that gives developers a fully managed infrastructure stack for running AI agents in the cloud. No containers to configure. No sandboxing to build. No state management to implement.
For SMBs building AI agents, this changes the calculus significantly. Here's what you need to know.
What Claude Managed Agents Actually Does
The pitch is straightforward: Anthropic handles the infrastructure, you define the task.
Specifically, the platform manages:
Secure execution
— Sandboxed containers with pre-installed packages (Python, Node.js, Go)
State management
— Checkpointing and persistent file systems across long-running sessions
Tool orchestration
— Built-in execution loops for bash commands, file operations, web search
Observability
— Session tracing, integration analytics, and troubleshooting in the Claude Console
Credentials and permissions
— Managed authentication for tools and external services
The key insight: developers were building this same stack over and over for every agent deployment. Anthropic productized it.
When This Matters for Your Business
Claude Managed Agents is designed for workloads that need:
Long-running execution
— Tasks that run for minutes or hours with multiple tool calls
Cloud infrastructure
— Secure containers without managing your own servers
Stateful sessions
— Persistent file systems and conversation history
Minimal infrastructure overhead
— No need to build your own agent loop or sandbox
For SMBs, the clearest use cases are:
Document processing agents that read, analyze, and extract information over hours
Research agents that search the web, compile findings, and generate reports
Coding agents that read codebases, make changes, and open pull requests
Workflow agents that connect to multiple tools and orchestrate multi-step processes
Real-World Deployments: What the Early Adopters Built
The product is already running in production at major companies:
Company
Use Case
Deployment Time
Notion
Custom Agents in private alpha — teams delegate work to Claude within workspaces
Active
Rakuten
Agents across product, sales, marketing, finance connecting to Slack and Teams
1 week per agent
Asana
AI Teammates — collaborative agents inside Asana projects
Accelerated development
Sentry
Debugging agent paired with Claude to write patches and open pull requests
Weeks, not months
Vibecode
Default integration for app deployment
Active
The Rakuten deployment is particularly notable. They built specialist agents for product, sales, marketing, and finance — each connecting to Slack and Teams — with deployments taking about a week per agent. Employees assign tasks and receive spreadsheets, slides, and apps as outputs.
This is the velocity that was previously reserved for companies with dedicated ML infrastructure teams.
Pricing: What It Costs
The pricing model is consumption-based:
$0.08 per session-hour
for active runtime
Standard Claude token rates
for input/output tokens
No fixed subscription or minimum commitment
For context: if your agent runs for 2 hours processing documents, that's $0.16 in infrastructure costs plus whatever tokens the agent consumes. A typical document processing workload with Claude Sonnet 4.6 might cost $0.50-$2.00 total depending on document complexity.
This is accessible pricing for SMBs. You're not building and maintaining a Kubernetes cluster — you're paying a few dollars for agent runtime that would have required weeks of engineering time to build yourself.
Performance: What the Data Shows
In internal testing on structured file generation tasks, Managed Agents improved task success by up to 10 percentage points compared to standard prompting loops. The largest gains came on harder tasks where the built-in orchestration and error recovery made the difference.
The platform also includes self-evaluation features (research preview) that allow Claude to iterate against defined success criteria. Multi-agent coordination, where agents can launch and direct other agents in parallel, is also in research preview.
How This Changes AI Strategy for SMBs
Before Managed Agents, deploying an AI agent required:
Building an agent loop (the orchestration layer)
Setting up secure execution (sandboxing, containers)
Implementing state management (database, file persistence)
Adding observability (logging, tracing, debugging)
Managing credentials and permissions
Handling errors and recovery
For an SMB without a dedicated infrastructure team, this was often a 4-8 week project just to get a basic agent into production.
With Managed Agents, you define the agent (model, prompt, tools), configure the environment (packages, network rules), and start a session. The infrastructure is handled.
The practical impact:
You can now deploy production AI agents in days rather than weeks, without infrastructure expertise.
What's Still in Research Preview
Two features are gated behind research preview access:
Multi-agent coordination
— Agents launching and directing other agents in parallel
Self-evaluation
— Claude iterating against defined success criteria
These features are likely to become more central as AI agents take on more complex workflows. If your use case involves multi-step processes with handoffs between agents, request early access.
Getting Started
Managed Agents is available now through the Anthropic API. Requirements:
A Claude API key
The beta header
managed-agents-2026-04-01
on all requests
Access is enabled by default for all API accounts
The
ant
CLI, launched alongside Managed Agents, provides a command-line client for managing API resources with YAML-based versioning.
The Bottom Line
Anthropic's move signals where the AI industry is heading: model providers are becoming platform providers. OpenAI has GPT-5.4 with desktop control. Google has Gemini in Workspace. Now Anthropic has Managed Agents.
For SMBs, the question shifts from "can we build the infrastructure?" to "what do we want our agents to do?"
The infrastructure problem isn't completely solved — you still need to design your agent, define its tools, and set appropriate guardrails. But the deployment complexity that killed so many agent projects just got significantly smaller.