2026-05-06

dreaming

Original source

Reading Log Article Claude Com Historical Backfill Agents Org Design LLM Research

Imported from historical reading log.

  • Claude Managed Agents update centered on three things: dreaming, outcomes, and multiagent orchestration.
  • Dreaming is a research-preview async job that reads an existing memory store plus past session transcripts and emits a cleaned/reorganized memory store with deduped facts, replaced stale entries, and new synthesized insights; original store remains unchanged.
  • Outcomes adds an explicit done target plus rubric-driven grading, turning a session from chat into iterative artifact production with a separate grader context feeding gap reports back to the agent.
  • Multiagent orchestration lets a coordinator agent delegate to specialized agents running in isolated persistent session threads while sharing the same container/filesystem.
  • Overall pattern: Anthropic is productizing more of the harness layer explicitly, memory maintenance, evaluator loops, and agent delegation, instead of treating them as app-side glue.
  • Strong newsletter angle: managed agents are becoming workflow infrastructure, not just a model wrapper.