Claude Managed Agents Public Beta: What's Actually Available vs What's Gated

Anthropic launched Claude Managed Agents into public beta earlier this year and the coverage has been enthusiastic - possibly too enthusiastic. Read past the launch blog posts and you'll find a consistent pattern: the features that make the product genuinely compelling are still gated behind a separate research preview application. What's actually available today is narrower and more infrastructure-focused than most headlines suggest.

This is not a takedown. The available feature set is genuinely useful for the right workloads. But the gap between what's marketed and what you can actually use is wide enough that it's worth mapping carefully before you build anything on top of it.

Last updated: June 10, 2026

---

What Managed Agents Delivers Today

The public beta ships four capabilities that work right now, without any waitlist.

Sandboxed execution environments. Every agent session runs inside an isolated container. Anthropic manages the default cloud sandbox, but teams can also point the environment config at self-hosted infrastructure. The ant CLI and REST API both support creating environments with configurable networking - unrestricted outbound by default, or locked down per your requirements. This is the genuinely useful core of the product: your agent can run bash commands, write files, and hit external APIs without those operations touching your application server.

Long-running sessions with persistence. Sessions stay alive across network disconnections and can be resumed. The event streaming model buffers events server-side so a dropped connection doesn't kill in-flight work. For tasks that take minutes to hours - code generation that iterates, data transformation pipelines, multi-step research - this is a meaningful improvement over stateless API calls wrapped in retry logic.

Tool execution via agent toolsets. The agent_toolset_20260401 tool type unlocks a pre-built set of tools: bash, file read/write, web search, and more. You declare them once when creating the agent definition and they're available to every session. The docs show the complete toolset, and individual tools can be scoped if you want to limit what a particular agent can do.

MCP server support. Agents can connect to MCP (Model Context Protocol) servers, which lets you wire in custom tools and data sources using the same protocol that Claude Code uses internally. If you've already built MCP integrations for your own Claude Code setup, those can be reused here.

The official quickstart on platform.claude.com walks through the full create-agent/create-environment/start-session flow in seven SDKs plus raw curl. All Managed Agents API requests require the managed-agents-2026-04-01 beta header - the SDK sets this automatically.

---

What's Still Locked

Two of the most-cited capabilities are not available in the public beta. They're in a separate "research preview" with gated access, meaning you have to apply and wait for approval.

Multi-agent coordination. Parallel task execution across multiple agents - the kind of fan-out that lets you decompose a big task into concurrent subtasks with results aggregated back - is research preview only. The public beta is strictly single-agent per session.

Self-evaluation loops. The ability for an agent to assess its own output, decide it's insufficient, and iterate without a human in the loop is also gated. What's available today is a single agent loop that goes idle when it decides the task is done. Self-evaluation and retry on quality criteria requires the research preview tier.

There's no published timeline for when these move to general availability. Anthropic's communications on this have been vague - "coming to more customers over time" is the level of specificity you're getting right now.

This matters because most of the compelling use cases in the launch coverage - autonomous research pipelines, self-correcting code generation, distributed agent teams - depend on one or both of these gated features.

---

Pricing Reality

The cost model has two components that stack on top of each other.

The $0.08/session-hour infrastructure cost (source) is billed for active agent runtime. If your session runs for 30 minutes, that's $0.04 in infrastructure on top of whatever you spent on tokens. For a high-volume pipeline running hundreds of sessions daily, this adds up quickly. A simple calculation: 500 sessions/day averaging 20 minutes each = 167 session-hours/day = $13.36/day in infrastructure costs alone, before tokens.

For bursty or low-volume workloads, the overhead is manageable. For continuous high-throughput pipelines, the cost math favors DIY infrastructure.

---

Lock-In Risk: Claude Only

This is the constraint that matters most for teams thinking long-term. Managed Agents supports Claude models only. There's no provider abstraction, no ability to route sessions to GPT, Gemini, or a local model. The agent definition model field takes a Claude model ID and only a Claude model ID.

If your evaluation determines that a different model performs better for a specific task six months from now, you're not swapping it in. You're rebuilding your agent infrastructure on a different platform. For teams that want to stay model-agnostic or hedge against provider pricing changes, this is a material concern.

---

What Teams Are Actually Shipping

Given the constraints, what's actually getting built with the available feature set?

The workloads that fit best are bounded, single-agent tasks where the value comes from reliable sandboxed execution rather than multi-agent coordination. Common patterns in the early adopter community include:

• Code generation with verification. Prompt the agent to write a script, execute it in the sandbox, return output. The sandbox execution is the value-add - you get confirmation the code actually ran, not just that it looks plausible.
• Document processing pipelines. Feed a batch of files into a session, extract structured data, write outputs. The persistence model handles large batches cleanly.
• CI/CD integration. Versioned agent definitions with environment promotion fit naturally into deployment pipelines. Teams are using this to run agents in staging vs production environments with different tool scopes.
• Structured data extraction. Anthropic's own testing reported task success improvements of up to 10 percentage points over standard prompting loops for structured file generation tasks.

None of these require multi-agent coordination or self-evaluation. They're useful, but they represent a fraction of the use cases the launch marketing suggested.

---

Open-Source Alternatives

If model flexibility or self-hosting matter to your team, three alternatives are worth evaluating honestly.

Multica is the closest open-source analog to Managed Agents. It supports multiple models, includes a task and team management UI, and is fully self-hostable. The gap is that it lacks container-level sandboxing and the credential vault isolation that Managed Agents provides.

Cabinet adds persistent agent memory and scheduled recurring tasks - capabilities Managed Agents doesn't currently offer at all. The tradeoff is no compute sandbox; Cabinet manages memory and scheduling but not execution isolation.

CrewAI is the most mature multi-agent framework of the three. Model-agnostic via LiteLLM, with a hosted management option and a large integration ecosystem. If you need multi-agent coordination today without waiting for research preview access, CrewAI is the practical path.

---

Make vs Buy Decision Framework

Use Managed Agents when:

• Your workload is single-agent, bounded tasks
• Sandboxed execution is the core requirement and you want managed infrastructure
• You're already committed to Claude and don't need model flexibility
• Compliance requirements benefit from Anthropic's enterprise data processing agreements
• CI/CD-style versioned agent deployments fit your workflow

Consider alternatives when:

• You need multi-agent coordination now, not on a waitlist
• Model-agnostic infrastructure is a requirement
• Cost at scale makes $0.08/session-hour prohibitive
• You want self-evaluation and autonomous iteration immediately
• Your team has the infrastructure capability to run sandboxed containers directly

The honest framing is this: Managed Agents is a good infrastructure product that happens to be marketed partly as a product it's not finished becoming yet. The sandboxing and session persistence are real and solid. The multi-agent orchestration is real but not available to most people reading this.

---

FAQ

Is Claude Managed Agents free to try?

No. There is no free tier for Claude Managed Agents. You need an Anthropic Console account and API key, and all usage is billed at standard token rates plus $0.08 per session-hour for infrastructure.

Can I use GPT-5 or Gemini with Managed Agents?

No. Claude Managed Agents supports Claude models only. There is no provider abstraction or model-swapping capability. If you need model flexibility, look at CrewAI or Multica instead.

What does "research preview" mean for multi-agent features?

It means gated access with a separate application process and no committed timeline for general availability. You cannot currently access multi-agent coordination or self-evaluation in the standard public beta - these require explicit approval from Anthropic.

How does the $0.08/session-hour cost work in practice?

It's billed on active session runtime, not token usage. A 30-minute session costs $0.04 in infrastructure fees plus whatever you spent on input and output tokens. Sessions that are idle but not terminated still accrue the session-hour charge.

Can I bring my own sandbox infrastructure?

Yes. The environment configuration supports self-hosted sandboxes via the type: self-hosted option in the environment config. This can reduce infrastructure costs if you already run container infrastructure, though you take on the operational burden.

Is Managed Agents suitable for production workloads today?

For single-agent, bounded tasks with sandboxed execution - yes. For multi-agent pipelines, autonomous self-evaluation, or model-agnostic requirements - not yet, or not without significant workarounds.

---

Official Sources

• Anthropic Managed Agents Quickstart - Official documentation with API reference and SDK examples
• Claude Managed Agents: Honest Review - Community analysis of available vs gated features and pricing breakdown