Anthropic Introduces Dynamic Workflows for Claude Agent Orchestration

At AI Engineer New York 2026, Anthropic introduced dynamic workflows (or "flows") for Claude Managed Agents, a new multi-agent orchestration system where Claude itself determines the organization of agents rather than relying on developer-configured rosters. The feature aims to solve the bottleneck developers face when managing numerous agents and PRs by allowing the lead agent to create plans and write orchestration programs that launch sub-agents and manage information flow.

How Dynamic Workflows Operate

Previously, Claude Managed Agents required developers to manually configure a coordinator and a roster of sub-agents. With dynamic workflows, Claude fully orchestrates its own multi-agent topology. The lead agent creates a plan, and Claude writes an orchestration program that launches agents, collects results, and manages interconnections. While Claude determines the division of labor, users can still pass in pre-configured specialist sub-agents if desired. The system is accessible via Claude Console, the interface for developers using the API.

Control, Budgets, and Optimization

During the Q&A, Anthropic engineers addressed concerns about control and cost. The lead agent does not know its token budget while running, but developers can set a session budget; engineers noted that informing models of budgets could make them "anxious." To improve efficiency over time, Anthropic highlighted a research preview feature called dreaming, which reorganizes memory stores by merging duplicates and replacing stale entries. While flows cost "much more" than single-agent approaches, they are described as "much more effective" for complex tasks.

Future Ambitions

Anthropic positions dynamic workflows as a way to "throw test-time compute at hard problems," aiming to make scientific discoveries and complex projects broadly doable by the general public. Although self-improving loops (such as a scanner recognizing its own inadequacy) are not built-in, engineers suggested combining hill-climbing and dreaming to achieve similar results. Sub-agents share a filesystem for communication, but direct API messaging between workflow sub-agents is not supported.