Claude Managed Agents Add Background Dreaming and Subagents

Anthropic introduced three core capabilities for Claude Managed Agents at its Code with Claude developer event on May 6, 2026. The release adds background memory consolidation, multiagent orchestration, and structured output grading to the platform. These updates shift the focus from single-prompt execution to continuous, self-improving background workflows supported by a new 300-megawatt compute partnership with SpaceX.

Background Memory Consolidation

The most significant architectural shift is the introduction of a “dreaming” process available in Research Preview. This capability allows agents to process session logs and consolidate context while idle. The system actively prunes stale information, merges duplicate entries, and resolves contradictory instructions in persistent files like CLAUDE.md.

Developers who add memory to AI agents typically rely on retrieval-augmented generation or manual file updates, but this built-in process automates state maintenance across active work sessions. You can configure the system to apply memory updates automatically or hold them for manual review before they affect the agent’s context window. It is optimized for long-running projects where retaining high-signal context is critical.

Independent Output Grading

The public beta of Outcomes provides a formalized framework for agent evaluation and error correction. Developers define success using specific rubrics. An independent grader model evaluates the output against these criteria in a separate, isolated context window to prevent reasoning bias. When the grader detects a failure, it pinpoints the specific issue, triggering the execution agent to attempt another pass automatically.

Internal testing indicates this iterative refinement process yields measurable gains over standard prompting strategies.

Task Category	Success Rate Improvement
Overall Task Success	+10.0 percentage points
Presentation Generation (.pptx)	+10.1 percentage points
Document Generation (.docx)	+8.4 percentage points

If you evaluate AI output in production applications, shifting to this built-in rubric system offloads the engineering overhead of building custom retry loops.

Parallel Execution and Capacity

The multiagent orchestration feature allows a primary lead agent to spawn and manage specialized sub-agents. A routing agent powered by Opus can parse a complex request and delegate execution steps to parallel Sonnet instances equipped with different tools and prompts. The sub-agents operate within isolated session threads but share a common filesystem, allowing the lead agent to aggregate their output files into a final response. Teams at Netflix utilize this architecture to parse massive parallel build logs, while legal firm Harvey uses it to draft long-form documents.

Anthropic also released webhooks to connect Managed Agents directly to external event triggers. To support these autonomous workloads, the company doubled the 5-hour rate limits for Claude Code across Pro, Max, and Team plans. They also removed peak-hour reductions for Pro and Max users. The infrastructure backing these sustained tasks includes a new partnership leveraging the SpaceX Colossus 1 data center, adding over 300 megawatts of compute capacity strictly for agent workloads.

If your application relies on multi-step generation, map out your workflow to isolate which components require deep reasoning and which need fast execution. You can route the complex planning to a lead Opus agent while assigning the parallel execution steps to specialized subagents in Claude Code to optimize both latency and token costs.