With Claude Managed Agents, Anthropic packs the infrastructure to run agents in production

Building AI agents has become a relatively straightforward endeavour – running them reliably in production, however, remains fraught with complexity.

While developers can stitch together models, tools, and prompts to get an agent working, turning that into something that can operate at scale – handling long-running tasks, managing memory, and staying within guardrails – often requires weeks or months of infrastructure work.

That’s the problem Anthropic is targeting with “Claude Managed Agents,” a new service that packages the underlying systems needed to deploy and run agents into a hosted platform.

Claude Managed Agents: ‘From prototype to launch in days’

With Claude Managed Agents, Anthropic is offering a managed service that handles the execution layer behind agent-based systems. Developers define what an agent should do – its tasks, tools, and constraints – while Anthropic runs it on its own infrastructure.

That includes support for long-running processes, access to external tools, and built-in memory, as well as sandboxed environments designed to keep agents within defined boundaries. The aim is to reduce the amount of engineering work required to move from a working prototype to something that can operate in production.

The company positions this as a shift away from assembling agent systems manually toward deploying them as services – much like the shift from self-hosted software to SaaS. So instead of wiring together orchestration logic, storage, and runtime environments, developers can rely on Anthropic to manage those layers behind the scenes.

“Until now, building agents meant spending development cycles on secure infrastructure, state management, permissioning, and reworking your agent loops for every model upgrade,” the company wrote in a blog post. “Managed Agents pairs an agent harness tuned for performance with production infrastructure to go from prototype to launch in days rather than months.”

Managed Agents is priced on a consumption basis. Standard Claude token rates apply, alongside an additional $0.08 per session-hour for active runtime, meaning costs are tied both to how much the model is used and how long agents are running.

Claude Managed Agents in action

A number of notable companies are already building agent-driven features on top of the system, often embedding them directly into existing products.

Notion, for example, is using Managed Agents to let teams delegate work to Claude inside its workspace, with multiple tasks running in parallel. Asana, meanwhile, has built what it calls “AI teammates,” agents that can pick up and complete assigned tasks within projects.

Elsewhere, Sentry is connecting its debugging tools to agents that can not only identify root causes but also write fixes and open pull requests.

A common thread permeates these various examples: agents are being positioned as part of day-to-day operations, handling defined tasks within existing systems in production environments. Put another way, Anthropic wants the world to know that Claude is ready for prime time.

Infrastructure becomes the product

The launch reflects a broader shift in how AI systems are being packaged. Much of the recent work has been on improving models or refining how agents reason through tasks. Here, Anthropic is focusing on what happens after that reasoning is complete.

Running agents in production introduces a different set of challenges: managing state, coordinating tasks over time, and ensuring systems behave predictably under load. Those are problems more familiar to infrastructure teams than prompt engineers.

That push toward managed environments is not happening in isolation. Anthropic has also recently moved to restrict the use of Claude subscriptions in third-party agent tools, citing infrastructure strain and terms-of-service violations, instead steering more complex or high-volume usage toward its own platform and API.

In that context, Managed Agents is part of an effort to bring agent workloads into environments that can be monitored, priced, and controlled internally. By wrapping those functions into a managed service, Anthropic is making a case that the next phase of agent development will depend on how easily they can be deployed and operated in production.

Whether that approach reduces the friction of building real-world agent systems, or simply shifts it into another layer of the stack, will depend on how much control developers are willing to hand over in exchange for convenience.