DEVOURED
Anthropic shipped 'dreaming' for Claude Managed Agents, where agents analyze their own past sessions to extract patterns and self-improve memory, with Harvey reporting 6x completion rate gains on legal document work.
What: On May 6, Anthropic launched three features for Claude Managed Agents: dreaming (scheduled analysis of past sessions to extract patterns and update memory automatically or via review), outcomes (separate grader evaluates output against rubrics and directs re-work until quality bar is met), and multiagent orchestration (lead agents delegate to specialist subagents running in parallel). Harvey reported 6x completion rate improvement with dreaming. Outcomes boosted task success by up to 10 points in internal tests. Netflix uses multiagent orchestration for parallel build log analysis, Spiral by Every for concurrent draft generation with quality enforcement, and Wisedocs for 50% faster document reviews.
Why it matters: This marks a shift from stateless AI to persistent systems that accumulate institutional knowledge across sessions. The two-tier architecture—memory for within-session learning, dreaming for cross-session pattern extraction—establishes a template for how agent systems build and refine knowledge over time, similar to how human teams develop expertise.
Deep dive
- Dreaming is a scheduled background process that reviews agent sessions and memory stores to extract recurring patterns, shared workflows across agents, team preferences, and restructures memory to keep it high-signal as it grows
- Control modes: dreaming can update memory automatically or require human review before changes land, giving developers flexibility based on trust level
- Outcomes let you define success as a rubric; a separate grader evaluates agent output in its own context window (isolated from agent reasoning to avoid confirmation bias) and pinpoints what needs fixing
- Self-correction loop: when output doesn't meet the rubric, the grader directs the agent to take another pass until quality bar is cleared
- Quality gains: outcomes improved task success by up to 10 points overall, +8.4% for docx generation, +10.1% for pptx in Anthropic's internal benchmarks, with largest gains on hardest problems
- Multiagent orchestration: lead agent breaks work into pieces and delegates to specialists with their own model, prompt, and tools running in parallel on shared filesystem
- Persistent events: agents can check back with each other mid-workflow because every agent remembers what it's done
- Full traceability: Claude Console shows which agent did what, in what order, and why across the entire delegation chain
- Webhooks: define an outcome, let agent run, get notified when complete—useful for long-running autonomous work
- Harvey uses dreaming to remember filetype workarounds and tool-specific patterns between sessions on legal drafting and document creation, driving ~6x completion rate increase in tests
- Netflix platform team built analysis agent that processes logs from hundreds of builds in parallel using multiagent orchestration, surfacing only recurring issues worth acting on
- Spiral by Every uses Haiku lead agent to field requests and delegate drafting to Opus subagents running in parallel, with outcomes enforcing editorial rubric and user voice pulled from memory—only drafts that clear the bar are returned
- Wisedocs uses outcomes to grade document reviews against internal guidelines, achieving 50% faster reviews while maintaining team standards alignment
- Availability: dreaming in research preview (request access), outcomes and multiagent orchestration in public beta as part of Managed Agents
Decoder
- Managed Agents: Anthropic's platform for deploying Claude-powered agents that persist across sessions with memory, tools, file access, and now cross-session learning via dreaming
- Dreaming: Anthropic-specific term for scheduled background analysis where agents review their own past sessions to extract patterns and curate memory for self-improvement between runs
- Grader: Separate evaluation instance in outcomes that checks agent output against rubrics in its own context window, isolated from the agent's reasoning to avoid the agent grading its own work
Original article
Claude Managed Agents launched features like dreaming, outcomes, and multiagent orchestration. Dreaming enhances agent improvement by analyzing past sessions to identify patterns, while outcomes allow agents to self-correct based on predefined success criteria. Multiagent orchestration optimizes complex task management by enabling agents to delegate tasks to specialized subagents, as utilized by companies like Harvey, Netflix, Spiral by Every, and Wisedocs.
Explore all 73 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi