Anthropic launched Dynamic Workflows for Claude Managed Agents in beta on October 9: Claude generates an orchestration program from the task, and a server executes it in the background, spawning sub-agents, passing intermediate results between stages and deciding when to branch or retry, so the main agent does not handle every hand-off. Per Anthropic's documentation, a single run can spawn at most 1,000 agents cumulatively, with 64 concurrent threads and a 24-hour default limit; at an AI Engineer New York 2026 workshop, Anthropic's CJ Avilla and Harrison Stall demoed a stock watchlist scan that split stocks across agents, checked results and produced a report. Anthropic says sub-agents cannot message each other directly and their threads are archived when a run ends, a completed status only means the program finished, and a session budget pauses new requests at the cap while in-flight requests still complete, so costs can exceed it; no built-in self-improvement is included.

Anthropic·Claude Managed Agents·Dynamic Workflows·+2