On April 8, 2026, Anthropic launched Claude Managed Agents in public beta, a hosted runtime for building and deploying cloud-based AI agents. Instead of assembling their own agent loop, sandbox, session storage and secret handling, developers configure an agent and let Anthropic run it, paying token rates plus $0.08 per active session-hour. The launch moved Anthropic from supplying models and SDKs toward operating the agent infrastructure itself, which changes the build-versus-buy calculation for teams that ship agents.
Key Facts#
- Launch: April 8, 2026, in public beta on the Claude Platform. API requests require the
managed-agents-2026-04-01beta header. - What it is: A hosted agent runtime and harness, described by Anthropic as a suite of composable APIs for building and deploying cloud-hosted agents at scale.
- Included: An orchestration harness with tool calling and error recovery, long-running sessions that survive disconnections, sandboxed code execution, credential vaults with scoped permissions, Model Context Protocol integration and skills.
- Pricing: Standard Claude Platform token rates plus $0.08 per active session-hour.
- Preview features: Multi-agent coordination launched as a research preview.
- Early customers: Anthropic named Notion, Rakuten, Asana, Vibecode, Sentry, Atlassian, General Legal and Blockit.
- Same-day tooling: Anthropic also released the
antCLI, a command-line client for the Claude API with YAML-based resource versioning.
What Happened#
Managed Agents packages the parts of an agent system that most teams end up building themselves. Anthropic's later technical write-up describes three core resources. An agent is a configuration of model, prompt, tools and guardrails. An environment is where work runs, including sandbox containers, networking rules and pre-installed packages. A session pairs an agent with an environment for one run, with its own isolated sandbox and a persistent event history.
The design separates what Anthropic calls "the brain from the hands." The reasoning harness runs apart from the code-execution sandboxes, and the two are connected through an append-only session log that records every model call, tool call and result. That log lets sessions resume after interruptions and lets idle containers be checkpointed without the developer running a database.
Credential handling is the most security-relevant part. Tokens for MCP servers, command-line tools and GitHub repositories live in a separate vault and are encrypted with envelope encryption. A proxy retrieves and decrypts them only on demand, so secrets never sit inside the sandbox where a prompt-injected agent could read them.
Customer quotes in the launch post focused on speed. Rakuten said it could deploy each specialist agent within a week, and Vibecode said its users could spin up agent infrastructure "at least 10x quicker." These are customer claims rather than independent measurements.
Background#
Through 2025, building a production agent meant stitching together a model API, an agent framework, a sandbox for running code, storage for conversation state, a secret store and monitoring. Anthropic had already released parts of that stack as software: the Claude Agent SDK packages the loop that powers Claude Code, and Agent Skills and MCP standardize procedures and tool access. Managed Agents takes the next step and operates those pieces as a service.
The release fits a broader industry pattern in which model providers compete on the runtime around the model, not just the model. For developers, that raises familiar platform questions about lock-in, data location and cost predictability.
Why It Matters for Developers#
For .NET teams, the decision is whether to host the agent loop yourself or rent it. The table below frames the trade-off at a high level.
| Concern | Claude Managed Agents | Self-hosted framework (for example, Microsoft Agent Framework) |
|---|---|---|
| Where the agent loop runs | Anthropic's platform | Your own services and infrastructure |
| Code execution sandbox | Provided, with self-hosted options added later | You choose and operate one |
| Session state and resumption | Built-in event log | You design storage and recovery |
| Secrets for tools | Managed vault outside the sandbox | Your secret store, such as Azure Key Vault |
| Model choice | Claude models | Depends on the framework and providers you use |
| Cost model | Tokens plus $0.08 per active session-hour | Tokens plus the infrastructure you run |
A managed runtime is attractive when time to production matters more than control, or when the team lacks experience running sandboxes safely. A self-hosted framework, covered in our Microsoft Agent Framework guide, fits better when agents must run inside your network, use several model providers or meet strict data residency rules. Our AI agent patterns guide covers the architectural choices on either path.
Whichever route you take, some practices carry over:
- Keep credentials out of the agent's reach. Anthropic's vault-and-proxy design is a good reference architecture for self-hosted systems too. Our secrets management guide covers the .NET equivalents.
- Log every step. An append-only record of model and tool calls is what makes agent behavior auditable and debuggable. The LLM observability and cost guide shows how to build that with OpenTelemetry in .NET.
- Budget runtime, not just tokens. At $0.08 per active session-hour, a thousand half-hour sessions add $40 of runtime on top of token costs. Long-idle sessions and runaway loops are the costs to watch.
What's Next#
Anthropic extended Managed Agents quickly after launch, according to its platform release notes. Built-in agent memory entered public beta on April 23, 2026. In early May, multi-agent orchestration and a feature called Outcomes moved to public beta, webhooks arrived for session and vault events, and a memory-consolidation feature called Dreams entered research preview. Self-hosted sandboxes and MCP tunnels, which let agents reach MCP servers inside private networks, followed on May 19. Managed Agents also became available through Claude Platform on AWS in May, scheduled deployments arrived in June, and in September 2026 permission policies gained an auto evaluation mode.
Open questions include how pricing evolves when the service leaves beta, how portable agent definitions are to other runtimes, and how enterprises will combine a hosted runtime with internal compliance controls. Self-hosted sandboxes and MCP tunnels suggest Anthropic is already adjusting the design for customers who need more control.