Building an AI agent is the easy part. Getting it into production with proper sandboxing, authentication, state management, and permissions? That’s where months disappear. Anthropic just took a sledgehammer to that problem.
The company launched Claude Managed Agents, a suite of composable APIs that lets teams build and deploy cloud-hosted agents at scale. It’s available now in public beta on the Claude Platform. And it signals a big strategic shift: Anthropic is moving from selling models to selling the entire agent runtime.
What happened
- Anthropic released Claude Managed Agents, a fully managed service for deploying production AI agents.
- You define your agent’s tasks, tools, and guardrails. Anthropic runs everything else on their infrastructure.
- A built-in orchestration harness handles tool calling, context management, and error recovery automatically.
- Pricing is consumption-based: standard Claude token rates plus $0.08 per session-hour for active runtime.
- Major companies like Notion, Rakuten, Asana, Sentry, and Atlassian are already building on it.
The numbers
- 10x faster time to production, according to Anthropic and multiple launch partners.
- Up to 10 points higher task success rate on structured file generation compared to standard prompting loops, with the biggest gains on the hardest problems.
- 3x faster development for one launch partner (a meeting prep agent that went from idea to shipping in days).
- Rakuten deployed each specialist agent within a week.
- Sentry shipped their code-fixing integration in weeks instead of months.
What you actually get
Here’s why this matters if you’ve ever tried to ship an agent to real users. Production agents need a long list of infrastructure that has nothing to do with the AI itself. Managed Agents bundles all of it into one service.
- Secure sandboxed execution. Code runs in isolated environments with authentication and tool execution handled for you. No more building your own container infrastructure.
- Long-running sessions. Agents can operate autonomously for hours. Progress and outputs persist even if the connection drops. This is table stakes for real work, and most DIY setups get it wrong.
- Multi-agent coordination (research preview). Agents can spin up and direct other agents to parallelize complex work. Think of a lead agent breaking a project into tasks and farming them out.
- Self-evaluation loops (research preview). You define outcomes and success criteria. Claude evaluates its own work and iterates until it meets the bar. This is where the 10-point task success improvement comes from.
- Scoped permissions and identity management. Agents get access to real systems with proper governance. Every tool call, decision, and failure mode is traceable in the Claude Console.
- Built-in observability. Session tracing, integration analytics, and troubleshooting guidance live directly in the console. You can inspect exactly what your agent did and why.
Who’s already using it
The launch partner list tells you exactly who Anthropic is targeting: enterprise teams that need agents embedded in existing workflows.
- Notion lets teams delegate work to Claude directly inside their workspace. Engineers ship code while knowledge workers produce websites and presentations. Dozens of tasks run in parallel.
- Rakuten deployed specialist agents across product, sales, marketing, finance, and HR that plug into Slack and Teams. Employees assign tasks and get back spreadsheets, slides, and apps.
- Asana built AI Teammates that work alongside humans inside projects, taking on tasks and drafting deliverables.
- Sentry paired their debugging agent Seer with a Claude-powered agent that writes patches and opens PRs. Developers go from a flagged bug to a reviewable fix in one flow.
- Atlassian is building agents for developers directly into Jira, so customers can assign tasks from workflows they already use.
- Vibecode uses Managed Agents as their default integration, letting users go from prompt to deployed app.
Why this is a big deal
This launch represents a fundamental shift in how Anthropic makes money and competes. Selling API tokens is a commodity business. Selling a managed agent runtime with built-in orchestration, governance, and observability is a platform business. The margins are better, the switching costs are higher, and the moat is deeper.
For enterprise buyers, the pitch is compelling. Instead of hiring a team to build agent infrastructure from scratch (sandboxing, state management, permissions, tracing), you get all of it out of the box for $0.08 per session-hour plus token costs. The ROI math works fast when you factor in the engineering salaries you’re not spending.
For the broader AI ecosystem, this is a signal that the “agent infrastructure” layer is consolidating. If Anthropic handles the runtime, teams build on Anthropic. That’s the playbook AWS ran with Lambda, and it worked.
5 things you should do right now
- Audit your current agent stack. If you’ve built custom infrastructure for sandboxing, sessions, or permissions, calculate what it costs you per month in engineering time. Compare that to $0.08 per session-hour.
- Pick one internal workflow to pilot. The fastest wins are repetitive, multi-step tasks: document processing, code review, data extraction. Start there.
- Request access to the research previews. Multi-agent coordination and self-evaluation loops are the most interesting features. Get in early.
- Read the docs before you commit. Managed Agents is in public beta. Understand the pricing model, limitations, and what’s still in preview versus what’s production-ready.
- Think about lock-in. Building on Managed Agents means building on Claude. If model portability matters to you, design your agent logic so the business rules live in your code, not in Anthropic’s orchestration layer.
The bottom line
Anthropic is no longer just selling a model. They’re selling the entire stack for getting agents into production. For teams that were already using Claude, this removes the biggest bottleneck (infrastructure) and lets them focus on the product. For teams evaluating AI platforms, the bar for what “production-ready” means just went up significantly.
The managed agent runtime is the new battleground. Anthropic just planted their flag.