Researched explainer, updated October 23, 2026. Product facts come from Anthropic’s announcements and documentation and from published reporting. We have not tested Claude Managed Agents ourselves, and we don’t score products.
Most agent updates this year have been about doing more things. Anthropic’s recent wave is more interesting because part of it is about agents getting better on their own. Alongside practical enterprise infrastructure, Anthropic introduced a technique it calls “dreaming” — and it’s one of the more thought-provoking ideas in agent development we’ve seen in 2026.
What’s new in Claude Managed Agents
Anthropic expanded its managed-agent offering on two fronts. On the infrastructure side, it added public-beta self-hosted sandboxes and a research-preview “MCP tunnels” feature. On the capability side, it expanded beta access to tools that let agents coordinate sub-agents and evaluate their own work using rubric-based outcomes — part of a broader push toward autonomous agents that handle long-running workflows in coding, finance, and law.
The “dreaming” technique, explained
This is the headline idea. “Dreaming” lets autonomous systems review their prior behavior between sessions, identify patterns, and improve future performance. Rather than starting each task cold, an agent can reflect on what it did before and carry forward lessons — loosely analogous to how sleep consolidates human learning.
If conventional agents are stateless contractors, a “dreaming” agent is one that actually learns from last week’s mistakes before showing up Monday.
It’s launching as a research preview, so this is early. But the direction matters: self-improvement between sessions is exactly the capability that separates a genuinely useful long-running agent from one that repeats the same errors indefinitely.
Self-hosted sandboxes & MCP tunnels
The infrastructure additions are aimed squarely at regulated enterprises. Self-hosted sandboxes let tool execution run on customer-managed or partner compute (like Cloudflare, Daytona, Modal, or Vercel) instead of the provider’s cloud. MCP tunnels let agents call internal MCP servers through an outbound-only encrypted gateway.
Together they solve a practical dilemma: how to use a managed agent orchestration layer while keeping sensitive data and credentials inside your own security perimeter. For banks, law firms, and healthcare — the exact sectors Anthropic is targeting — that’s often the difference between “interesting demo” and “approved for production.”
Pros & cons
Strengths
- ‘Dreaming’ enables genuine cross-session learning
- Self-hosted sandboxes keep data in your perimeter
- MCP tunnels protect internal credentials
- Sub-agent coordination for complex workflows
- Rubric-based self-evaluation of outputs
- Aimed at high-value coding, finance, legal work
Limitations
- ‘Dreaming’ is a research preview — very early
- Several features still in public beta
- Real setup complexity for the security features
- Best value needs in-house engineering maturity
Who it’s for
It’s for engineering teams building long-running, autonomous agents — especially in regulated industries that need execution and data to stay inside their own perimeter. Hold off if you just need a simple task bot; the power here is in infrastructure and learning that most lightweight use cases won’t tap.
How it compares
| Capability | Claude Managed Agents | OpenAI Frontier | Gemini / Antigravity |
|---|---|---|---|
| Cross-session learning | “Dreaming” (preview) | Optimization loops | Limited |
| Self-hosted execution | Yes (beta) | Limited | Partial |
| Sub-agent orchestration | Yes | Yes | Yes (Antigravity) |
| Best fit | Regulated, long-running | Large enterprise | Google ecosystem |
Anthropic’s distinctive angle is the combination of a security-first execution model with an early but real bet on agents that improve themselves. Frontier is broader for enterprise management; Google’s Antigravity is strongest inside its own stack.
Frequently asked questions
What does ‘dreaming’ actually do?
What problem do self-hosted sandboxes solve?
What are MCP tunnels?
Is this production-ready?
What has changed since launch (October 2026)
- Anthropic announced dreaming on May 6, 2026, at its Code with Claude conference, as a research preview that developers must request access to.
- Dreaming runs between sessions, reviews up to 100 earlier sessions, and proposes memory updates that teams can approve or reject; the model itself is not retrained.
