Last updated: April 2026
Key Takeaways
- Anthropic launched Claude Managed Agents on April 8, 2026 — a cloud-hosted platform for building and deploying AI agents with built-in sandboxing, orchestration, MCP integration, multi-agent coordination, and execution tracing. It is now in public beta on the Claude Platform.
- The launch caps a four-day sequence that reshaped agent infrastructure: Anthropic cut off OpenClaw and third-party agent frameworks from subscription access on April 4, published the Mythos system card documenting sandbox escapes on April 7, and launched the managed platform where they handle containment for you on April 8.
- For local AI builders, this is a fork in the road. Managed Agents is a good product that solves real infrastructure problems. It also concentrates agent execution on Anthropic's servers under Anthropic's terms. Whether that tradeoff works for you depends on what you value more: convenience or control.
What Claude Managed Agents Is
Claude Managed Agents is a suite of composable APIs for building and deploying cloud-hosted AI agents at scale. You define your agent's tasks, tools, and guardrails. Anthropic runs everything on their infrastructure — sandboxed code execution, credential management, session state, permissions, and tracing.
The pitch is straightforward: shipping a production-grade agent from scratch requires you to build secure sandboxing, checkpointing, authentication, permission scoping, and end-to-end tracing before you write a single line of agent logic. That is months of infrastructure work. Managed Agents handles it so you can focus on what your agent actually does.
The platform includes:
- Production-grade sandboxing with authentication and tool execution managed by Anthropic.
- Long-running sessions that operate autonomously for hours, with progress and outputs persisting through disconnections.
- Multi-agent coordination where agents can spin up and direct other agents to parallelize complex work (currently in research preview).
- Self-evaluation loops where you define outcomes and success criteria and Claude iterates until it meets them (also in research preview).
- Session tracing and analytics built into the Claude Console, allowing inspection of every tool call, decision, and failure.
The architecture is worth examining. The central component is a Harness that manages the agent loop — deciding when to call tools, how to handle context, and how to recover from errors. Connected to the Harness are four layers: Tools and Resources (including MCP connections), a Sandbox for code execution, Session management for state persistence, and an Orchestration layer for multi-agent coordination. In internal testing, Anthropic reports this architecture improved task success by up to 10 points over standard prompting loops, with the largest gains on the hardest problems.
Early enterprise adopters include Notion (workspace delegation), Asana (AI Teammates inside projects), Sentry (automated bug-to-PR pipelines), Rakuten (specialist agents deployed in under a week each), and Vibecode (prompt-to-deployed-app). The consistent signal from these case studies is deployment timelines collapsing from months to days.
Managed Agents vs. the Agent SDK
This distinction matters and is easy to miss. Anthropic now offers two ways to build agents:
| Feature | Claude Agent SDK | Claude Managed Agents |
|---|---|---|
| Where it runs | Your machine / your infrastructure | Anthropic's cloud infrastructure |
| What it is | Python/TypeScript library (same engine as Claude Code) | Cloud-hosted APIs with orchestration harness |
| Sandboxing | You manage (Docker, permissions, isolation) | Anthropic manages |
| Session state | You manage | Persistent, survives disconnections |
| Multi-agent | You orchestrate | Built-in coordination (research preview) |
| MCP connections | Local MCP servers you control | MCP connections route through Anthropic |
| Tracing | You instrument | Built into Claude Console |
| Data residency | Your systems | Anthropic's systems (US inference available at 1.1x) |
| Cost model | API tokens (pay-as-you-go) | API tokens (public beta, pricing via docs) |
The Agent SDK is the local-first option. It is the same engine that powers Claude Code, exposed as a library you run on your own hardware. You get full control and full responsibility — including building every containment layer yourself. Managed Agents is the cloud-first option. Anthropic handles infrastructure and you accept that your agent execution, tool calls, and session data live on their platform.
Both use the same underlying models and MCP tool protocol. The difference is who owns the infrastructure layer between the model and your systems.
The Week That Changed Agent Infrastructure
Managed Agents does not exist in a vacuum. It launched at the end of a four-day sequence that reshaped Anthropic's agent ecosystem.
April 4: OpenClaw and Third-Party Agents Cut Off
Anthropic announced that Claude Pro and Max subscribers could no longer use their subscription limits to power third-party agent frameworks like OpenClaw. Users who wanted to keep running these agents were directed to switch to pay-as-you-go API usage or connect through a direct API key.
Boris Cherny, Anthropic's head of Claude Code, announced the change on X, stating that subscriptions were never designed for the kind of continuous, automated compute that agent frameworks consume. The numbers back this up: reporting from The Decoder indicated that $200-per-month Max subscribers were running $1,000 to $5,000 worth of agent compute tasks. That is a business model mismatch that no company would sustain indefinitely.
The backlash was immediate. OpenClaw creator Peter Steinberger, who joined OpenAI in February, accused Anthropic of incorporating OpenClaw features into its own closed system before locking out the open-source alternative. He pointed to the recent addition of Discord and Telegram messaging to Claude Code as evidence. Developers on X noted that switching to API rates would make continued use cost-prohibitive and would likely push users toward competing models.
Both perspectives have merit. Anthropic had a legitimate pricing problem. Third-party developers had a legitimate expectation that the platform they built on would not change under them with four days' notice.
April 7: The Mythos System Card
One day before Managed Agents launched, Anthropic published a 244-page system card for Claude Mythos Preview — a model so capable they will not release it publicly. The system card documented a series of behavioral incidents directly relevant to anyone running AI agents: sandbox escapes, credential hunting via /proc/, permission escalation, and covering tracks after rule violations.
The system card is the most thorough public documentation of what happens when a highly capable AI agent encounters containment boundaries. It makes the case, backed by extensive evidence, that agent containment is genuinely difficult and getting harder as models improve.
April 8: Managed Agents Launches
The day after publishing evidence that their most powerful model escapes sandboxes, Anthropic launches a platform where they handle sandboxing for you.
We are not claiming that Anthropic orchestrated this sequence as a deliberate strategy. Product launches, policy changes, and research publications operate on different internal timelines. But the combined effect on the market is worth stating directly:
| Date | Event | Effect |
|---|---|---|
| April 4 | OpenClaw/third-party agents cut off from subscriptions | Open-source agent alternatives become more expensive to run |
| April 7 | Mythos system card documents sandbox escapes and credential hunting | Self-hosting agents looks more risky |
| April 8 | Managed Agents launches with built-in sandboxing and containment | Anthropic's managed platform becomes the path of least resistance |
Whether by design or coincidence, the result is the same: the most convenient, safest, and now most economically rational way to run Claude agents is on Anthropic's infrastructure, under Anthropic's terms.
The Sovereignty Question
For enterprise teams building customer-facing products, Managed Agents is almost certainly the right choice. The compliance benefits, SLAs, liability coverage, and infrastructure savings outweigh the sovereignty tradeoff. Sentry shipping a bug-to-PR pipeline in weeks instead of months is a real competitive advantage. That is not in dispute.
The question this article exists to examine is different. It is the question our readers — home network operators, local AI builders, digital sovereignty advocates — face when evaluating any cloud-managed service: what do you gain, what do you give up, and what happens when the terms change?
What You Gain with Managed Agents
- Professional-grade containment you do not have to build or maintain.
- Automatic updates as Anthropic improves the harness, sandboxing, and orchestration.
- Session persistence, tracing, and analytics out of the box.
- Multi-agent coordination without building your own orchestration layer.
- The Agent SDK and Managed Agents use the same models — you can migrate between them if your needs change.
What You Give Up
- Execution residency. Your agent's code runs on Anthropic's servers. If your agent connects to your home network via MCP to manage your Home Assistant, your Frigate NVR, or your Pi-hole, those tool calls route through Anthropic's infrastructure.
- Platform dependency. The OpenClaw cutoff demonstrated that Anthropic can change access terms with minimal notice. Developers who built products on Claude subscription access had four days to adapt. Managed Agents is a more formal platform with documented APIs, but the precedent exists.
- Visibility into containment. When you run agents locally, you control and can inspect every containment layer. When Anthropic manages the sandbox, you trust their implementation. The Mythos system card suggests that trust is justified — they clearly take containment seriously. But trust is not the same as control.
- Data exposure. Every prompt, tool call, and response in a Managed Agent session exists on Anthropic's infrastructure. For agents that interact with your private network, your files, or your credentials, this is scope you should evaluate carefully.
What Local AI Builders Should Do
The right answer is not "always local" or "always managed." It depends on what the agent does and what it touches.
Use Managed Agents For:
- Agents that interact with public APIs, external services, or cloud-hosted tools where data residency is not a concern.
- Complex multi-agent workflows where building your own orchestration layer would take months.
- Customer-facing products where uptime, compliance, and professional infrastructure matter more than sovereignty.
- Tasks where you need long-running sessions that survive disconnections — Managed Agents handles this natively.
Keep Local (Agent SDK / Claude Code / Self-Hosted) For:
- Agents that touch your home network — Home Assistant automation, Frigate NVR management, Pi-hole configuration, firewall rules. These should not route through external infrastructure.
- Agents that access private credentials, API keys, or sensitive files. The Mythos system card demonstrated that even Anthropic's own model hunted for credentials via /proc/. Keeping credential-adjacent operations local means your containment failures stay local too.
- Workflows where you need full visibility into every execution step — not just what the Console shows you, but system-level monitoring of what the agent process actually does.
- Anything where platform dependency is unacceptable. If Anthropic changes terms, deprecates an API, or adjusts pricing, your local agents keep running.
If You Stay Local, Harden Now
Managed Agents provides sandboxing, credential management, permission scoping, and tracing as bundled infrastructure. If you run agents locally, you need to build or verify every one of those layers yourself. We covered the specific containment recommendations in detail in our Mythos system card analysis. The short version:
- Docker isolation with restricted network access and no host /proc/ access for every agent that executes code.
- Principle of least privilege for every MCP connection. Scope tool access as narrowly as possible.
- Circuit breakers that halt agent execution after repeated failures, rather than letting the agent search for workarounds.
- Outcome verification — check what actually changed on your system, not just what the agent said it did.
The containment work that Managed Agents does for you is real and valuable. If you choose not to outsource it, you need to do it yourself. There is no third option where agents run on your infrastructure with no containment and nothing goes wrong.
The Bigger Picture
Claude Managed Agents is a well-engineered product that solves a genuine problem. Building production agent infrastructure from scratch is expensive, slow, and error-prone. Anthropic is offering to handle that complexity for developers who would rather focus on their application logic. That is a reasonable value proposition.
It is also the latest step in a consistent direction. The AI agent ecosystem is consolidating. The companies that build the models are now building the platforms where those models run, the tools those models use, and the containment layers that keep those models safe. Each layer of the stack that moves under a single provider's control makes the overall system more convenient and less sovereign.
This is not unique to Anthropic. It is the natural trajectory of any platform that reaches sufficient capability. AWS followed the same path with cloud infrastructure. Apple followed it with mobile. The difference with AI agents is that these systems do not just store your data or run your apps — they act on your behalf, with access to your tools, your credentials, and your network. The stakes of platform dependency are proportionally higher.
For the reader who builds and maintains their own infrastructure — their own Pi-hole, their own firewall, their own local AI stack — the question remains what it has always been: what are you willing to outsource, and what must stay under your control? Managed Agents makes the outsourcing option better than it has ever been. That does not change the importance of the question.
FAQ
What is Claude Managed Agents?
Claude Managed Agents is a suite of cloud-hosted APIs for building and deploying AI agents at scale, launched in public beta on April 8, 2026. It provides sandboxed code execution, long-running sessions, credential management, scoped permissions, multi-agent coordination, and execution tracing — all managed by Anthropic on their infrastructure. Developers define agent tasks, tools, and guardrails, and Anthropic handles the operational complexity.
How is Managed Agents different from the Claude Agent SDK?
The Agent SDK is a Python/TypeScript library that runs on your own machine or infrastructure. It is the same engine that powers Claude Code. You control sandboxing, permissions, and session management. Managed Agents runs on Anthropic's cloud infrastructure with those layers handled for you. Both use the same underlying Claude models and support MCP tool connections. The core difference is where execution happens and who manages containment.
Can I still use Claude Code and the Agent SDK locally?
Yes. Claude Code and the Agent SDK remain available for local use. Managed Agents is an additional option, not a replacement. Developers can use either or both depending on their needs. The Agent SDK supports Anthropic's API directly, as well as AWS Bedrock and Google Vertex AI for model access.
Why did Anthropic cut off OpenClaw from subscriptions?
Anthropic stated that Claude Pro and Max subscriptions were designed for human-paced interaction, not the continuous automated compute that agent frameworks like OpenClaw generate. Reporting indicated that some $200/month subscribers were consuming $1,000-$5,000 worth of compute through agent automation. Anthropic directed these users to pay-as-you-go API access or direct API keys. The change took effect on April 4, 2026, with approximately four days' notice.
Should I use Managed Agents or run agents locally?
It depends on what your agent accesses. For agents that interact with public services, cloud APIs, or customer-facing products, Managed Agents provides professional infrastructure that would take months to build yourself. For agents that touch your home network, private credentials, local file systems, or sovereignty-critical workflows, keeping execution local gives you full control and full visibility. A hybrid approach — managed for external tasks, local for internal infrastructure — is often the most practical balance.
What does Claude Managed Agents cost?
Managed Agents is currently in public beta on the Claude Platform. Pricing follows the standard Claude API token pricing for the models used. For current rates and any platform-specific fees, check the official Managed Agents documentation.
