Claude Managed Agents Add Dreaming, Outcomes and Multiagent Orchestration

What Anthropic released on 6 May 2026

On 6 May 2026 at the Code w/ Claude event in San Francisco Anthropic introduced four new features for Claude Managed Agents. The platform has been on the market only since 8 April 2026, just over four weeks. With the new features the competition shifts from pure model quality toward agent engineering and self-improvement.

• Four features: Dreaming as research preview, Outcomes and Multiagent Orchestration in public beta, Webhooks for task notifications
• Introduced by Ami Vora (Chief Product Officer) and Boris Cherny (creator of Claude Code) at the Code w/ Claude event
• Announced alongside: doubled rate limits for Claude Code Pro and Max, peak-hour throttling removed
• SpaceX partnership for the Colossus data centre with more than 300 megawatts of additional capacity and over 220,000 GPUs within one month
• API volume has grown 17 times year on year, per Anthropic

Dreaming: How agents learn between sessions

Dreaming is a scheduled background process that analyses past agent sessions, identifies recurring patterns and curates the memory store. Anthropic likens it to the consolidation of hippocampal memories during sleep. Concretely Dreaming reads up to 100 past sessions and the existing memory store, removes duplicates and outdated entries and rebuilds the memory while preserving the original transcripts.

• Detects recurring mistakes, converging workflows and team-wide preferences that a single agent cannot see on its own
• Updates either automatic or only after manual review
• Supported models: Claude Opus 4.7 and Sonnet 4.6, billed via standard API tokens, no additional licence fee
• Harvey case: the legal AI firm reports roughly six times higher completion rates through institutional knowledge from Dreaming, for example file-handling protocols
• Status: research preview, access only through Anthropic's request form

The idea of letting agents review their own sessions is not new. The Hermes approach from Nous Research proposed a similar self-improvement concept in the open-source space. Anthropic's variant has the advantage that it requires no custom training and integrates into an existing platform.

Outcomes: Rubrics instead of prompts

Outcomes shifts the responsibility for quality from the prompt engineer to the grader. Developers define a rubric that describes what success looks like. A separate grader agent in its own context window evaluates the result against those criteria. If the result falls short, the grader identifies the gaps and the agent revises up to three times, in exceptional cases up to 20.

Outcomes is public beta and can be activated through an API header. Companies with a measurable definition of success can put the mechanism into production today, without new models or new SDKs.

Multiagent Orchestration: Up to 20 specialists under a lead agent

Multiagent Orchestration solves the context window problem for large tasks. A lead agent decomposes the work and delegates each piece to a subagent with its own model, prompt and toolset. The specialists run in parallel on a shared filesystem and contribute their results back into the lead agent's context.

• Up to 20 distinct agent roles, up to 25 parallel threads
• Single-level model: no sub-orchestrators, no recursive delegation
• Full traceability through the Anthropic Console

In the Anthropic demo a lead agent investigates an incident while subagents fan out in parallel through deploy history, error logs, metrics and support tickets. Netflix uses the pattern in production for build logs of hundreds of pipelines. Spiral by Every combines Haiku 4.5 as a specialist with Opus 4.7 as the lead and reports frontier-level quality at roughly one fifth of the cost.

European perspective: Data residency and GDPR

Claude Managed Agents run only on Anthropic's own infrastructure in the United States. For European mid-market and enterprise customers with GDPR obligations, direct usage is therefore only partially possible. Anthropic opened a Munich office in March 2026 and hired a Head of DACH and CEE, but data residency in the US is unchanged.

Bedrock AgentCore adds 3 to 4 seconds of latency because requests flow through AWS-owned endpoints. For many back-office workflows that is acceptable, for real-time customer interactions usually not. The EU AI Act high-risk deadlines take effect on 2 August 2026 and address exactly the accountability question in multi-agent setups that the text does not yet make explicit.

Challenges and risks

More autonomy means greater damage when decisions are wrong. Hallucinations in agents are no longer just wrong answers, they become wrong actions with real consequences. Multi-agent setups amplify these risks because errors cascade across agent boundaries.

• Memory poisoning: Malicious entries in the memory store persist across sessions and shape future behaviour. A known attack scenario: prepared court filings that a legal agent later retrieves and acts on.
• Cascading failures: A failing subagent can compromise the entire workflow, the lead agent does not always recognise this.
• Audit gap with Dreaming: When memory is automatically curated between sessions, it is unclear how a later explanation of an agent decision can be reconstructed in practice.
• Regulatory uncertainty: The EU AI Act does not define agentic systems explicitly, the distributed responsibility across multiple agents is legally unresolved.
• Identity problem: Multi-agent setups demand an identity model for non-human actors that most companies still lack.

What companies should do now

The new features are only worth it if architecture and governance grow with them. Companies should not skip the platform question out of enthusiasm for Dreaming and Multiagent Orchestration. Clarify both in parallel.

1. Data residency before feature choice

   Companies under GDPR should start directly with AWS Bedrock AgentCore in Frankfurt. Allow direct Anthropic usage only for prototypes with non-personal data.

2. Introduce Outcomes before Dreaming

   Rubrics force the team to a measurable definition of success before agents learn from it. Without clear success criteria, Dreaming learns the wrong patterns.

3. Start multiagent setups small

   Two or three specialists under one lead agent, not 20 at once. Add further subagents only after four to six weeks of stable operation.

4. Define memory governance

   Who may create, modify or delete entries in the memory store? How are source, timestamp and trigger of every entry logged? Without these answers Dreaming remains an audit risk.

5. Test audit requirements

   Take a sample of agent decisions and fully reproduce and trace them before going live. That includes the curated memory contents, the grader rubric and each subagent's tools.

6. Keep the vendor strategy open

   Anthropic's features are an advantage today; in three to six months similar functions are expected from OpenAI, Google and Salesforce. Design the architecture portably, for example with Model Context Protocol as the tool layer.

Treat enterprise agent governance as the leverage, not the feature set. A well-defined Outcomes rubric is worth more than three additional subagents.

Further Reading

Frequently Asked Questions