Anthropic introduces dynamic workflows for Claude Managed Agents to coordinate multi-agent orchestration

At AI Engineer New York 2026, Anthropic described how its new dynamic workflows feature—announced on Friday—enables Claude Managed Agents to autonomously orchestrate multi-agent topologies for complex tasks. The feature, referred to as 'flows,' allows Claude to create an orchestration program that launches agents, collects results, and manages information flow without requiring developers to manually configure coordinators or rosters.

Previously, developers had to configure a coordinator and a roster of subagents for multi-agent tasks. With dynamic workflows, Claude itself determines how to divide and orchestrate work across discrete, verifiable units, while still permitting developers to introduce pre-configured specialist sub-agents if desired. All functionality is accessible via Claude Console, the interface for developers using Claude through API.

During a demonstration, Anthropic showed a stock watchlist scanner workflow that illustrated a scan phase distributing individual tickers to isolated-context agents and a review phase compiling findings into a report. The presenters emphasized that flows are designed for 'your most ambitious workloads,' though the example was a simpler illustration of the pattern.

Audience questions highlighted practical concerns, including how Claude decides the number of agents a task requires. Anthropic responded that this is governed by evaluations and lower defaults, with developers able to explicitly encourage greater parallelization for ambitious tasks. Concerns about token budget control were addressed by noting that developers can set a session budget, though the lead agent does not have real-time budget awareness.

Visibility into sub-agent actions was addressed by confirming that developers can inspect payloads and transitions through graphing functionality in Claude Console, enabling deep understanding of inter-agent communication and task progression.

Anthropic acknowledged that flows cost significantly more than single-agent approaches but argued they are more effective for complex tasks. The company framed flows as a way to 'throw test-time compute at hard problems' and enable general users to tackle ambitious projects, including large scientific research, though specific capabilities for such projects were not detailed.

During Q&A, Anthropic suggested that features like 'dreaming'—a research preview feature that reorganizes memory stores—could help improve workflow efficiency over time, though automatic self-improvement of flows was not built in. Subagents share a filesystem for information exchange but cannot directly message each other via API.