Anthropic demos workflows to run 1,000 Claude agents

Anthropic demos workflows to run 1,000 Claude agents

Anthropic staffers demonstrated how Claude Managed Agents can coordinate thousands of nested subagents without hand-built orchestration systems, showing workflows with a single event stream, token budgets and live steering.

CJ Avilla and Harrison Stall of AI Engineer presented the flows as an approach to handling harder tasks through nested agents working in parallel. Stall said a single agent "keels over much, much earlier" while dynamic workflows allow arbitrarily complex arrangements of parallel agents.

The demonstration addressed problems with hand-rolled orchestration, which often fails on container restarts and leaves users with opaque graphs. The session showed a single event stream alongside controls for token budgets and live steering.

According to Claude Code Docs, a dynamic workflow is a JavaScript script written by Claude and executed in the background by a runtime. The approach advances beyond the single-coordinator roster used in Managed Agents V1 from May.

To frame the coordination challenge, the presenters drew on software development history: Ken Thompson's 1969 one-month Unix build, IBM's 1,000-engineer OS/360 effort and Linus Torvalds' 2002 move from a 36-actor star to a layered trust network after Linux reached scale. That shift in organization provided context for the session's focus on nested agents rather than a single coordinator.

Read the source