Anthropic Unveils Dynamic Workflows for Complex Engineering Tasks in Claude Code

Anthropic has introduced dynamic workflows in Claude Code, a new capability designed to handle complex, long-running engineering tasks by orchestrating tens to hundreds of parallel subagents. This feature bridges the gap between single-agent interactions and custom-built agent teams, allowing the AI to plan, break down tasks, and execute them in parallel with built-in verification steps. Currently in research preview, the tool is available across the Claude Code CLI, Desktop, and VS Code extension, as well as via API on platforms including Amazon Bedrock, Vertex AI, and Microsoft Foundry.

The system is engineered to address problems too large for a single agent pass, such as bug hunts across entire services or migrations touching hundreds of files. Dynamic workflows consume substantially more tokens than typical Claude Code sessions, prompting Anthropic to recommend starting with scoped tasks to gauge usage. The feature requires auto mode to be enabled for the best experience, and the coordination mechanism operates outside the main conversation thread to maintain plan integrity during long-running tasks.

Early access users within Anthropic have reported success in tasks such as identifying dead code and refactoring large codebases. Developer Jarred Sumner utilised dynamic workflows to port the Bun runtime from Zig to Rust, completing the task in eleven days with 99.8% of the existing test suite passing. The workflow mapped Rust lifetimes for every struct field, wrote behavior-identical ports for hundreds of files, and ran a fix loop until builds and tests were clean.

Availability varies by plan: it is on by default for Max and Team plans and API users, but off by default for Enterprise plans, where administrators must enable it via managed settings. The system includes a confirmation step before the first workflow triggers, allowing users to review the planned execution. Progress is saved during runs, enabling interrupted jobs to resume from where they left off rather than restarting.

Anthropic states that the technology aims to reduce tasks that previously took weeks or quarters into days. Agents address problems from independent angles, with other agents attempting to refute findings until answers converge. This capability represents a significant shift in how large-scale software engineering challenges are approached, offering end-to-end handling with verification steps before results are presented to the user.