Anthropic · 2026-05-28 · major
Anthropic Ships Dynamic Workflows in Claude Code — Tens to Hundreds of Subagents Run in Parallel, Klarna and CyberAgent Already On, Powered the Bun Zig-to-Rust Port
Claude Code now plans, fans out, and verifies long-running work with up to hundreds of parallel subagents in one session. Available on Max, Team, and Enterprise plans, in CLI, Desktop, VS Code, and via the API.
Claude Code learns to fan out a single task across hundreds of subagents, then check their work before merging the result.
Key specs
| Subagents per session | tens to hundreds |
|---|---|
| Early customers | Klarna, CyberAgent |
| Availability | Max / Team / Enterprise |
What is it?
A new orchestration layer inside Claude Code that breaks large tasks into subtasks, dispatches independent subagents to work on them in parallel, then runs an independent verification pass before integrating results. Interrupted jobs resume from where they stopped rather than restarting.
How does it work?
Users either request workflows explicitly or enable the `ultracode` setting so Claude Code triggers them automatically. The orchestrator runs across CLI, Desktop, the VS Code extension, the API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Each subagent has its own context and the parent checks outputs for regressions before they land.
Why does it matter?
It's the first time Claude Code formalizes long-running, multi-hour migrations and audits as a first-class primitive. The post cites the Bun Zig-to-Rust port (960K lines, six days) as a real use case. The catch: dynamic workflows can consume substantially more tokens than a normal session, so Anthropic recommends starting scoped.
Who is it for?
Engineering teams running codebase-wide migrations, bug hunts, or audits that previously needed weeks of human time.
Try it
Toggle `ultracode` in settings or ask Claude Code for a workflow on a scoped task.