Anthropic on May 7, 2026 expands its Claude Managed Agents suite with three features — Dreaming, Outcomes

Anthropic announced on May 7, 2026 that it has added three capabilities to its Claude Managed Agents platform intended to make agents more reliable in production. The release follows the Managed Agents platform launch in April and was reported by Matthias Bastian. The new additions — Dreaming, Outcomes and Multiagent Orchestration — arrive as a mix of research preview and public betas.

Dreaming is an asynchronous review job that scans an existing memory store and, optionally, up to 100 past sessions to detect patterns, remove duplicate or outdated entries, and produce a consolidated, organized memory artifact while leaving the original memory intact. The feature supports Claude Opus 4.7 and Claude Sonnet 4.6, is billed under standard API token usage, and is controllable from the Claude Console where developers select a memory store, pick a model, specify sessions and launch the job.

Outcomes adds a formal, rubric — based evaluation flow. Developers provide a document of concrete success criteria and a separate evaluator — the grader — checks agent outputs against that rubric in its own isolated context window so it cannot be influenced by the agent’s internal chain of reasoning. If an output fails the rubric, the grader reports the gaps and the agent can revise by default up to three times; this retry behavior is configurable up to a ceiling of 20 attempts.

Multiagent Orchestration centralizes coordination by running a lead agent, or coordinator, that delegates tasks to specialized sub-agents. Each sub-agent runs in its own thread with an isolated context, its own model, system prompt and dedicated tools, while sharing a common file system. The system supports up to 20 distinct agents and can execute as many as 25 threads concurrently, enabling parallel workflows such as simultaneous code review and test generation.

Together, the three features address practical needs for builders deploying agents: Dreaming targets cross — session learning and memory hygiene by distilling insights without altering original records; Outcomes provides an auditable correctness layer that separates judgment from generation; and Multiagent Orchestration allows teams to decompose complex tasks into isolated, specialized units that can work in parallel. Anthropic frames these additions as steps toward more adaptive, production — ready agents.

Availability is specific: Dreaming is available as a research preview with access via a request form on the Claude website, while Outcomes, Multiagent Orchestration and Memory are listed as public betas within the Managed Agents platform. Developers are directed to the Claude Console, the product documentation and the company product blog for implementation details and integration guidance.