Anthropic launches dynamic workflows, letting Claude orchestrate multi-agent topologies on its own

Dynamic workflows are a new feature in Claude Managed Agents, positioned as "a new kind of multi-agent orchestration built for the largest workloads." Previously developers had to manually configure a coordinator and a set of sub-agents in Claude Console; now the lead agent first makes a plan, then writes an orchestration program that launches agents, collects results, and manages information flow on its own, so developers no longer manage connections between stages. Preset specialized sub-agents can still be passed in via agent configuration if needed. Anthropic demoed with a stock watchlist scan: the scanning phase distributes tickers to context-isolated agents, and the review phase aggregates results into a report. Cost and control are the main trade-offs. Anthropic technical staffer Harrison Stall said workflows are "much" more expensive than single-agent approaches but "much more effective" at finding every detail. The model doesn't know its budget at runtime, so developers can only set a session budget; simple tasks don't parallelize heavily by default and need explicit encouragement to spawn more sub-agents. Sub-agents exchange information through a shared file system, but the API can't directly instruct one workflow sub-agent to message another; developers can use Claude Console's graphical features to inspect the actual payload the lead agent passes to sub-agents. For developers automating with the Claude API, this hands orchestration to the model and suits long-running, splittable, verifiable tasks; for small teams watching costs, single-agent with manual orchestration remains cheaper. Anthropic also suggests pairing it with the research preview feature dreaming to tidy memory storage, so repeatedly run flows gradually reduce overhead.