Claude Code now spins up hundreds of parallel subagents in one session

Anthropic added dynamic workflows to Claude Code, a feature that lets Claude plan a large engineering task, fan it out across many parallel subagents in a single session, check the intermediate results, and save progress as it goes. The pitch is that work that used to take weeks of careful coordination can run unattended, with the model writing its own orchestration scripts and verifying results before they come back to you.

The concrete example Anthropic leads with is Bun's port from Zig to Rust. Jarred Sumner, the project's founder, says he used dynamic workflows to translate 750,000 lines of code in 11 days, with 99.8% of the test suite passing. Other suggested uses are codebase-wide bug hunts, security audits, and large migrations across thousands of files: the kinds of jobs where the cost of a wrong answer is high, so the model's adversarial self-check matters more than raw speed.

The feature is available on Max, Team, and Enterprise plans. Anthropic positions it as the bridge between Claude used interactively in a terminal and the long-running agent work that has shown up in research demos for over a year without yet becoming part of how most teams ship code.

Why it matters

If you have been using Claude Code interactively, the question to ask is which of your migrations, audits, or refactors you have been postponing because they would take days to babysit. Those are the workloads this is built for. The Bun port number is striking, but run a small workflow on your own codebase first and read the diffs carefully before trusting unattended jobs to land on main.