Claude Managed Agents: AI Agents That Improve While They Sleep

Imagine your AI assistant not only remembers what you told it today but sleeps on it overnight and comes back better tomorrow. That is exactly what Anthropic's new Dreaming feature does for Claude Managed Agents.

On May 6, Anthropic unveiled three new capabilities at its 'Code with Claude 2026' conference that fundamentally change how AI agents work.

What are Claude Managed Agents?

Managed Agents are cloud-hosted AI agents that autonomously handle multi-step tasks — not just answering a question but executing entire workflows. They have been available in public beta since April 2026.

The three new features in detail:

1. Dreaming — AI Learns in Its Sleep

The concept is inspired by human memory consolidation: just as our brains process the day's experiences during sleep and decide what goes into long-term memory, a 'dreaming' agent reviews past sessions and extracts patterns.

- Recurring mistakes are identified and corrected
- Proven workflows are saved and reused
- Duplicates in the knowledge base are merged
- Outdated information is replaced with current data

The result: the agent improves with every session without manual retraining. You control whether changes are applied automatically or reviewed first.

2. Outcomes — Making Success Measurable

With Outcomes, you define upfront what a good result looks like. A separate grader evaluates the output in its own context window and sends the agent back until quality meets your criteria. According to Anthropic's internal benchmarks, this improves task success by up to 10 points — especially on the hardest problems.

3. Multiagent Orchestration — Teamwork for AI

A lead agent can delegate tasks to specialist subagents that work in parallel on a shared filesystem. Each subagent has its own model, prompt, and tools. The entire flow is traceable in the Claude Console.

What does it cost?

- Tokens: Standard API pricing with no markup
- Runtime: 0.08 USD per session-hour (only active runtime counts, idle time is excluded)
- Dreaming: Standard token costs for the selected model
- Pay-as-you-go billing — no fixed subscription required

Real-world examples:

- Harvey (Legal AI): The legal software Harvey uses Dreaming so agents remember file-type workarounds and tool-specific patterns between sessions. Completion rates increased roughly sixfold
- Netflix: Already deployed multiagent orchestration for its platform team

Additional updates:

- Self-hosted Sandboxes (Public Beta): Run tool execution on your own infrastructure — via Cloudflare, Daytona, Modal, or Vercel
- MCP Tunnels (Research Preview): Secure connections to private MCP servers without public endpoints

What this means for you:

Dreaming is currently in research preview and primarily targets developers building custom agents via the API. But for end users, it signals where AI is heading: agents that self-improve, form teams, and autonomously verify quality.

Three things you can do now:

- For developers: Open the Claude Console and test your first Managed Agent with Outcomes — the public beta is available to all API accounts
- For everyone: Keep the concept of 'Dreaming' in mind — when your AI tool suddenly gets much better, this technique might be behind it
- Privacy note: Dreaming stores insights from past sessions. Consider what data you entrust to a long-term learning agent

Sources: claude.com/blog/new-in-claude-managed-agents, 9to5mac.com/2026/05/07/anthropic-updates-claude-managed-agents-with-three-new-features, siliconangle.com/2026/05/06/anthropic-letting-claude-agents-dream-dont-sleep-job, sdtimes.com/ai/new-in-claude-managed-agents-dreaming-outcomes-and-multiagent-orchestration