Introducing dynamic workflows in Claude Code

Anthropic announced dynamic workflows for Claude Code, positioned as a new working mode that lets Claude handle longer, more complex engineering tasks. Unlike a single agent replying all at once, dynamic workflows plan dynamically according to the user's prompt, break the task into subtasks, launch dozens to hundreds of subagents in parallel, and within the same session have them search, modify, verify, and push back on each other's results separately, before finally consolidating into a single coordinated answer. Anthropic emphasizes that this kind of flow is suited to tasks that a single agent struggles to complete in one pass, such as a bug hunt in a large legacy codebase, a security audit spanning an entire service, a profiler-driven performance review, framework replacement and API-deprecation migration across hundreds or thousands of files, and plans that require multi-angle stress testing before formal adoption. The article also mentions the case of Bun's port from Zig to Rust: Jarred Sumner used dynamic workflows to produce roughly 750,000 lines of Rust, getting the existing test suite to a 99.8% pass rate, taking about 11 days from the first commit to merge; however, the article also explicitly notes that this port has not yet entered production. The feature is currently a research preview, supporting the Claude Code CLI, Desktop, and the VS Code extension, aimed at Max, Team, and Enterprise plan users, and also usable via the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Ways to activate it include directly asking Claude to build a workflow, or turning on Claude Code's dedicated ultracode setting to let Claude decide on its own when to use a workflow. It is worth noting that dynamic workflows consume more usage than a regular Claude Code session, and on the first trigger will display what is about to run and request confirmation; enterprise administrators can also disable it via managed settings. For Taiwanese development teams, this means AI coding tools are advancing from "assisting with single-step modifications" toward "coordinating long-running, multi-agent, recoverable engineering tasks," but cost, auditability, and production-adoption risk still need conservative evaluation.