Building a production AI agent used to require weeks of infrastructure work before you could write a single line of business logic. Sandboxing, credential handling, state management, execution tracing — all of it landed on your engineering team’s plate before the agent could do anything useful.

Anthropic’s answer, launched in public beta on April 8, is Claude Managed Agents: a hosted infrastructure service that handles the entire operational layer, so teams can focus on what the agent actually does rather than the machinery keeping it running.

What It Is

Managed Agents is a fully managed harness for running Claude as an autonomous agent. You define the tasks, tools, and guardrails. Anthropic handles everything else: sandboxed code execution, authentication, checkpointing, credential management, and execution tracing.

The service supports long-running sessions — agents can operate continuously for hours and preserve progress through disconnections. This is a meaningful departure from standard API calls, which time out and lose state. For workflows involving research, document processing, or multi-step business operations, persistent sessions change what is realistically automatable.

Access is through the API with a managed-agents-2026-04-01 beta header. The design is deliberately infrastructure-agnostic, built to accommodate future harness types and sandbox configurations as Claude’s capabilities evolve.

The Pricing Math

The service runs at $0.08 per session-hour, on top of standard Claude API token rates. An agent running continuously around the clock costs roughly $58 per month in runtime charges before token costs — comparable to a basic SaaS subscription rather than a significant infrastructure line item.

For a one-hour task run daily, the runtime charge is about $2.40 per month. For a department deploying multiple agents handling real workloads, the cost structure is dramatically cheaper than the engineering time previously required to build equivalent infrastructure from scratch.

Who Is Already Using It

Four enterprise customers confirmed public beta deployments at launch, each in a different operational context.

Notion is using Managed Agents to let teams delegate tasks — coding, slide creation, spreadsheet work — without leaving their workspace. Multiple agents run in parallel on separate tasks simultaneously.

Rakuten deployed specialist agents across product, sales, marketing, finance, and HR. Each went live in under a week, which Rakuten noted was far faster than their previous internal agent deployments.

Asana built AI Teammates that sit inside project workflows and pick up tasks directly when assigned, just like a human team member would.

Sentry built an agent that runs end-to-end from flagged bug to open pull request — fully autonomous, without human intervention in the middle.

These are not prototype demos. They are production deployments handling real workflows inside real organisations, which is the practical signal that matters here.

Why This Is Different From What Came Before

Several agent infrastructure options existed before this launch. Amazon Bedrock Agents, Azure AI Agent Service, Google Vertex AI Agent Builder, LangChain, and CrewAI all compete for enterprise agent infrastructure spend. The distinction Anthropic is drawing is tight integration between the model and the managed environment.

Because Claude and the agent harness come from the same provider, Anthropic can make architectural decisions — like the external context object that lives outside Claude’s context window and persists durably in the session log — that are difficult to replicate when the model and infrastructure are separate products.

The TechRadar coverage described the launch as promising “10x faster” time to production. Anthropic’s own framing is more measured but directionally consistent: prototype to launch in days rather than months.

What This Means for Business

The practical shift here is that the barrier to building a production AI agent just dropped significantly. Not in a theoretical sense — in the immediate, operational sense that a non-enterprise-scale engineering team can now deploy autonomous agents handling real business processes without building the underlying infrastructure first.

This matters most for two kinds of businesses:

Teams that have already identified workflows to automate but stalled on implementation complexity. The weeks spent on sandboxing, auth, and state management before getting to business logic was a real filter that eliminated otherwise viable projects. That filter is largely gone now.

Teams evaluating whether AI agents are realistic for their size. The Rakuten deployment — multiple specialist agents across departments, each live in under a week — demonstrates that this is not exclusively an enterprise-with-a-large-platform-team story. The managed infrastructure removes the scale requirement for the initial deployment.

The counterpoint worth noting: removing infrastructure friction also removes the forcing function that caused teams to think carefully about agent scope, failure modes, and oversight before shipping. The governance layer that slowed deployment was not entirely waste. Moving fast on agent deployment without that overhead requires deliberate attention to what happens when an agent does something unexpected.

The businesses getting the most from this shift will be those that combine the speed Managed Agents enables with the operational thinking to define clear scope, meaningful guardrails, and monitoring that tells you when an agent is drifting from its intended behaviour. That is not a reason to avoid AI agents — it is a reason to approach deployment deliberately rather than experimentally.

---

If you’re deciding where to start with agents, start here. The free Working With Claude field guide walks through the ecosystem, Claude Code, and a real rollout plan. Get your copy.