CJ Avilla and Harrison Stall of AI Engineer demoed flows to run thousands of nested Claude subagents without hand-built harnesses.
They framed the problem with Ken Thompson's 1969 one-month Unix build, IBM's 1,000-engineer OS/360 effort and Linus Torvalds' 2002 shift from a 36-actor star to a layered trust network after Linux hit scale.
Stall said a single agent "keels over much, much earlier" while dynamic workflows enable arbitrarily complex nested parallel agents for harder tasks.
The new session shows a single event stream, token budgets and live steering, a fix for hand-rolled orchestration that often fails on container restarts and opaque graphs.
Per Claude Code Docs, a dynamic workflow is a JavaScript script that Claude writes and a runtime executes in the background, an advance over the single-coordinator roster in Managed Agents V1 from May.
