Anthropic adds dynamic workflows to Claude Code

Anthropic on May 28 said it is introducing dynamic workflows in Claude Code, a new mode designed to handle large engineering tasks end-to-end. The company said the feature lets Claude plan work dynamically, split it into subtasks, and run tens to hundreds of subagents in parallel before returning a single coordinated result.

The feature is aimed at problems that are difficult to finish in one pass with a single agent. Anthropic pointed to examples such as bug hunts across an entire service, migrations that touch hundreds or thousands of files, and plans that need to be tested from multiple angles before a team commits to them.

According to Anthropic, dynamic workflows are available now in research preview in the Claude Code CLI, Desktop app, and the VS Code extension for Max, Team, and Enterprise plans if administrators enable it. The company said the feature is also available through the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Anthropic said users can start a workflow in two ways. They can ask Claude directly to create one, or they can turn on a new Claude Code setting called ultracode, which is accessible through the effort menu and sets effort to xhigh while allowing Claude to decide when a workflow should be used.

The company said the first time a workflow triggers, Claude Code shows what is about to run and asks for confirmation. Anthropic also said organization admins can disable workflows through managed settings, and that Enterprise plans have dynamic workflows turned off by default at launch.

Anthropic warned that dynamic workflows can use substantially more tokens than a normal Claude Code session. Because of that, the company recommended starting with a scoped task to understand usage before trying larger jobs.

In a section on early use, Anthropic said early access users and teams inside the company have applied the feature to codebase-wide bug hunts, profiler-guided optimization audits, and security audits. The company said the same pattern can also be used for hardening passes, including checks for authentication, input validation, and unsafe patterns across a codebase.

Anthropic also said dynamic workflows can support large migrations and modernization work, including framework swaps, API deprecations, and language ports across thousands of files. For critical work, the company said workflows can run independent attempts on the same problem and use adversarial agents to try to break the result before it is shown to the user.

To illustrate the approach, Anthropic pointed to a recent rewrite of Bun. The company said Jarred Sumner used dynamic workflows to port Bun from Zig to Rust, with 99.8% of the existing test suite passing, roughly 750,000 lines of Rust produced, and eleven days from first commit to merge.

Anthropic said one workflow mapped the correct Rust lifetime for every struct field in the Zig codebase. Another wrote each .rs file as a behavior-identical port of its .zig counterpart, with hundreds of agents working in parallel and two reviewers on each file.

The company said a fix loop then drove the build and test suite until both ran clean. After the port landed, Anthropic said an overnight workflow addressed unnecessary data copies and opened a pull request for each item for final review. Anthropic said all of that was handled by dynamic workflows, although it added that the feature is not yet in production.

Under the hood, Anthropic said Claude begins by planning from the prompt, breaking work into subtasks, and fanning those tasks out across parallel subagents. Results are checked before they are folded back into the final answer, and the run keeps iterating until the outputs converge.

The company said this design is meant for parallel and long-running work that can take hours or days. It also said progress is saved as the run continues, so interrupted jobs can resume instead of restarting from the beginning.