Anthropic's Claude Managed Agents add dynamic workflows for self-orchestrating agents
At the AI Engineer conference in New York, Anthropic technical staff CJ Avilla and Harrison Stall presented dynamic workflows ("flows"), a feature announced the previous Friday as part of Claude Managed Agents that lets Claude decide and orchestrate its entire multi-agent topology: a lead agent makes a plan, writes an orchestration program, launches subagents and collects results, so developers no longer configure a coordinator and roster by hand. Flows are available in Claude Console; in a stock-watchlist demo they cost "much more" than a single-agent approach but were better at finding all the details, and developers can set a session budget while the model itself does not know its budget at runtime. The speakers also pointed to the "dreaming" research preview for reorganizing memory stores, noted that subagents share a filesystem but cannot message each other through the API, and said flows do not self-improve on their own.
It hands multi-agent orchestration decisions to the model itself, a step toward Anthropic's goal of long-running, ambitious workloads.