Anthropic's Managed Agents Blitz: Is Hosted Agent Infrastructure the New Walled Garden?

In April 2026, Anthropic did something that looked subtle but wasn't. It stopped selling you a model and started selling you a place to live.

Claude Managed Agents launched in public beta on April 8. On the surface, it's a hosted runtime for AI agents. You define a config, spin up a sandbox, start a session, and Claude runs autonomously for minutes or hours while your app streams results. No agent loop to write. No sandbox to provision. No error recovery to wire up. Anthropic handles the plumbing.

The pitch is compelling if you're trying to ship. Enterprises that spent months building agent infrastructure can now deploy in days. Notion, Rakuten, Asana, and Sentry had production workloads running before the public beta even opened. Anthropic's internal benchmarks showed a 10-point improvement in task success rates over standard prompting.

But underneath the convenience sits a strategy that should look familiar to anyone who lived through the cloud infrastructure wars of the 2010s. Anthropic doesn't just want to sell you Claude. It wants to host the agent loop, persist the session state, manage the sandbox, and become the floor your entire agent architecture stands on. Once you're there, leaving is not a migration. It's a rebuild.

What Managed Agents actually is

The architecture is well-designed. Anthropic's engineering team split the agent runtime into three decoupled components: a Session (append-only event log), a Harness (stateless loop that calls Claude and routes tool calls), and a Sandbox (isolated container where code runs). Each can fail and restart independently. If the harness crashes, a new one picks up the session log and resumes. It's the "pets vs. cattle" model from DevOps applied to agent infrastructure, and it's smart.

Pricing is hybrid. Standard Claude token rates plus $0.08 per session-hour when agents are actively running. A session handling 10,000 support tickets might cost somewhere around $37, depending on how many steps each task takes and which model you're using. It's not cheap, but it's not the cost that should worry you.

The thing to watch is where the state lives. Session data sits in Anthropic's database. Your agent's memory, its conversation history, its sandbox filesystem state. All of it stored server-side by Anthropic. You can delete sessions and files through the API, but you cannot export the agent's runtime state in a format another platform could ingest and continue from. The session log is append-only and Anthropic-specific. There is no "download my agent and run it on OpenAI" button, and there won't be.

Is this the AWS playbook, replayed?

This is not a novel move. It's the same strategy AWS deployed against on-premise data centres between 2006 and 2016.

AWS made cloud infrastructure trivially easy to adopt. Spin up an EC2 instance in seconds. Store terabytes in S3 without buying a SAN. Scale with a few API calls. The value was undeniable, and the switching costs arrived later, quietly, in the form of proprietary services. DynamoDB isn't portable. Lambda isn't portable. The data egress fees are punitive by design. Organisations that went all-in on AWS in 2012 found themselves locked in by 2018, not because anyone tricked them, but because AWS made the first step so easy that the architecture hardened around it before anyone noticed.

Anthropic is doing the same thing at the agent layer. Managed Agents makes the first step trivial. Define an agent. Start a session. Ship. No infrastructure team required. But every session log you accumulate, every agent workflow you build against Anthropic's proprietary harness, every sandbox configuration you tune for Claude's specific tool-calling behaviour, becomes a reason you can't leave.

And this isn't theoretical. Anthropic followed up the Managed Agents launch in May 2026 with Claude Platform on AWS: IAM-native authentication, CloudTrail auditing, billing through AWS Marketplace. The same procurement channel enterprises already use for everything else. The same "it's already in your AWS bill, just click here" friction removal that made cloud lock-in so effective the first time around. Anthropic is not just building walled gardens. It's building them inside the walls you already live in.

What you give up when someone else runs the loop

The word "managed" does a lot of euphemistic heavy lifting. Here is what it actually means in practice.

Your agent's decision loop runs inside Anthropic's infrastructure. You don't see the intermediate reasoning unless Anthropic exposes it. You don't control when compaction happens, how the harness prioritises tool calls, or what the error-recovery logic looks like when a sandbox times out. The loop is Anthropic-proprietary. If it changes behaviour between releases, you adapt. If Anthropic deprecates a tool or modifies the sandbox environment, your agents change with it.

The model lock-in is absolute. Managed Agents runs Claude and only Claude. There is no multi-model routing, no fallback to GPT or Gemini, no ability to swap the underlying model if Claude's pricing changes or its performance degrades on your specific workload. The Claude Agent SDK supports multi-model setups. Managed Agents does not. That's the trade: convenience for portability.

And then there's the pricing opacity. Hybrid billing (tokens plus runtime hours) makes cost modelling harder than pure token-based pricing. You're paying for the model and the infrastructure together, in proportions that shift depending on task complexity and session duration. A price change to either component, or a modification to how "active running" is measured, ripples through your costs in ways that are hard to predict.

The portable alternative is already here

You don't have to build directly on top of any model provider's infrastructure. The no-code and low-code ecosystem has been solving this abstraction problem for years, just at a different layer.

Platforms built on model-agnostic architecture, like Stacker, let you wire agents through an abstraction layer that treats the model as a configurable component rather than a foundation. You build your workflows, your permissions model, your data architecture once, and the platform handles which model executes which task. If Claude gets expensive next quarter, you swap. If Gemini ships a feature that matters to your use case, you route certain tasks to it. The agent logic, the session management, the user-facing interface, they all live in your platform, not in Anthropic's database.

This isn't about avoiding Claude. Claude is excellent. I use it. You probably should too. The question is whether it should own the infrastructure layer beneath your agents or just be one engine you plug in.

The practical difference shows up fast. A Managed Agents session is a black box Anthropic operates. A model-agnostic platform gives you the same agent capabilities (sandboxed execution, state persistence, tool routing) but keeps the orchestration logic in your stack, where you can inspect it, modify it, and point it at different models when your needs change. The convenience gap between the two approaches has been shrinking. The portability gap hasn't.

How to evaluate agent infrastructure decisions

If you're weighing up whether to build on Managed Agents or a portable platform, I'd suggest four questions.

Whose database does my agent's memory live in? If the answer is "the model provider's," you should know what happens to that data when you want to change providers. The answer, right now, for Managed Agents: it stays there. You can delete it. You can't migrate it.

Can I swap the model without rewriting the agent? If the answer is no, you're not building with agents. You're building with Claude. That might be fine for now. Just be honest about what "now" means to your roadmap.

Who decides when the agent loop changes? On Managed Agents, Anthropic does. The harness is their code. When they update it, your agents update. If you're fine with that, great. If your use case requires deterministic behaviour or auditable decision paths, you need a loop you control.

What does my migration path look like in 18 months? Not "will I migrate." Just "could I, if I needed to." If the answer involves rebuilding from scratch, the architecture has already made the decision for you.

Anthropic is building something useful with Managed Agents. The engineering is solid, the developer experience is clean, and the speed-to-production advantage is real. But the strategy behind it is not neutral. It is designed to make starting easy and leaving hard, same as every infrastructure-as-a-service play that came before it. Recognise the pattern, and build accordingly.