Claude Managed Agents: Anthropic brings three new features Apfelpatient

Anthropic is sharpening the profile of its cloud agent platform: Dreaming, Outcomes, and Multi-Agent Orchestration are three new features designed to make Claude-based agents more adaptive, goal-oriented, and better suited for teamwork. Major customers like Netflix are already using the new tools in production.

With Claude Managed Agents, Anthropic launched a key building block for AI agents in the cloud in April – a platform that significantly reduces the technical effort required to build and operate such agents. In parallel, Anthropic also completely rebuilt its Claude Code experience for Mac, adding features like Routines. The three new features now introduced for Claude Managed Agents demonstrate the platform's future direction: it is designed not only to execute tasks but also to learn from them, independently monitor goal achievement, and collaborate with specialized sub-agents in complex workflows.

Dreaming: How Claude's agents learn between sessions

The first new feature, called Dreaming, has launched as a Research Preview. It extends Claude's existing memory capabilities with a scheduled process that scans past sessions and memory repositories, extracts patterns, and curates stored memories. The goal is for agents to improve over time without requiring manual intervention from developers.

The level of user control is configurable. Dreaming can either automatically update the memory or present suggestions for review before changes are applied. Anthropic clearly describes its interaction with the existing memory function: Memory records what an agent learns while working – Dreaming refines these memories between sessions, compares insights across multiple agents, and keeps the memory up to date.

Outcomes: Success criteria as a separate evaluation mechanism

The second feature addresses a classic problem in agent workflows – the question of when a task is truly complete. With Outcomes, users can define an evaluation matrix that precisely describes what constitutes a successful outcome. The agent then works specifically towards meeting these criteria.

The actual evaluation is performed by a separate grader running in its own context window. This ensures that the test is not influenced by the agent's reasoning. If something is amiss, the grader identifies the specific vulnerability, and the agent initiates another pass. Additionally, an outcome can be defined, an agent can be run independently, and the result can be reported via webhook – a workflow that is crucial for asynchronous tasks.

Multi-agent orchestration: Lead agent coordinates specialists

Anthropic takes its most ambitious step with multi-agent orchestration. A lead agent breaks down a complex task into sub-tasks and delegates each sub-task to a specialized sub-agent with its own model, prompt, and tools. Anthropic describes an investigation as an example where the lead agent conducts research while sub-agents simultaneously sift through deployment history, error logs, metrics, and support tickets.

The specialists work on a shared file system and return their results to the overall context of the lead agent. Since all events are persistently stored, the lead agent can access interim results from the sub-agents at any time and flexibly adjust the coordination. According to Anthropic, Netflix is already using this feature productively for its platform team.

What is the significance of the update?

With Dreaming, Outcomes, and Multi-Agent Orchestration, Anthropic addresses three of the most pressing challenges of productive agent systems: the ability to learn across sessions, clear performance measurement, and the distribution of complex tasks across specialized units. This further positions the company as a specialist in agent-based AI – an area that has long since become a strategic focus in international competition.

For Anthropic, the update is more than just a technical detail. It marks the transition from a platform that automates individual tasks to an infrastructure on which companies can map entire parts of their operations in the form of AI agents. The fact that Netflix is already using multi-agent orchestration in production demonstrates how quickly such systems are becoming established in the enterprise sector.

Anthropic's agent strategy is taking clear form

With these three new features, Claude Managed Agents becomes one of the most mature cloud agent platforms on the market. Mac users will likely experience the effect indirectly – through improved responses, more adaptive workflows, and more stable automations. How quickly these new features gain traction beyond the Research Preview should become clear in the coming weeks. (Image: Shutterstock / Stockinq)