Anthropic's workshop moves from coordinating company-research sessions to scheduling recurring work. A deployment supplies an agent definition, schedule, initial messages and optional resources. Memory stores remain separate resources that a builder attaches to sessions. The live demonstration encounters failed runs, so the segment does not establish that every scheduled execution succeeds.
The presenters preview Dreaming, which analyzes a chosen period of session transcripts and updates a memory store with lessons from repeated problems. Directly modifying skills and prompts is described as future work. They emphasize task-specific evaluation; audience questions highlight the risk that observed patterns are not sufficient evidence of improved outcomes. Dreaming is an experimental research preview, not guaranteed autonomous self-improvement.
The Q&A distinguishes the managed tool loop from an application's broader orchestration, explains event-based observability and describes workspace separation. Fine-grained role-based access, graph-based orchestration UI and an on-premises offering are not shown as currently available. The API product supports customer-facing applications and team agents, while personal use remains a different workflow.
Watch on YouTube
