Anthropic's dynamic workflows let Claude run its own agents

Anthropic framed dynamic workflows around a much older idea. Avilla and Stall opened by invoking Ken Thompson building the original Unix operating system and Linux evolving from one person, Linus Torvalds, to layers of trusted maintainers. Their point was that agents need the same kind of organization for the ambitious projects Anthropic envisions for flows.

The feature grew out of a "harness for every task" concept inside Claude Code. What changed is who does the organizing. Claude Managed Agents previously required developers to configure a coordinator and a roster of subagents; with dynamic workflows, Claude writes an orchestration program that launches agents, collects results, and manages information flow. Users can still pass in pre-configured sub-agents through the agent config if they want specialists.

In a demo, a stock watchlist scanner showed the pattern: a scan phase distributed individual tickers to agents with isolated contexts, and a later review phase examined the findings before producing a report. Avilla said flows cost much more than a single-agent approach but were much more effective at finding all of the details. Audience questions pressed on control — how many agents a task needs (Stall pointed to evals and lower defaults), budget visibility, and whether subagents can message each other (they share a filesystem instead). For cheaper repeated flows, Avilla pointed to dreaming, a research preview feature that reads an existing memory store alongside past session transcripts and produces a reorganized one. Asked whether flows could power large scientific research, Stall would not detail how Anthropic computes such problems but said enabling general users to tackle them is the goal.