Anthropic Dynamic Workflows in Claude Code: What It Is and Why It Matters

On May 28, 2026, alongside Claude Opus 4.8, Anthropic shipped dynamic workflows in Claude Code. The one-line version: Claude writes a JavaScript script that orchestrates a swarm of subagents to do a large task, running in the background while your session stays responsive. It is a research preview, and it is a bigger deal than the headline agent count suggests.

What it actually is

Per Anthropic's documentation, a dynamic workflow is a JavaScript script that orchestrates subagents at scale. You describe a task; Claude writes the script; a runtime executes it. The clever part is where the intermediate results live: in the script's variables, not in Claude's context window. That is what lets a single run chew through an enormous job without blowing up the context.

The numbers that matter

• Up to 16 agents run concurrently (fewer on low-CPU machines).
• 1,000 agents total per run — a hard cap Anthropic added specifically to stop runaway loops.
• Anthropic's blog frames the scale as "tens to hundreds of parallel subagents in a single session."

Adversarial verification, not just parallelism

The differentiator is not the agent count — it is that agents can adversarially review each other's findings before anything is reported. The bundled /deep-research workflow cross-checks sources, votes on each claim, and drops the ones that do not survive the cross-check. That verification loop is what turns a pile of parallel output into something you can actually trust.

How you turn it on

• Include the word "workflow" in a prompt (Claude highlights it; alt+w to ignore).
• /effort ultracode — combines xhigh reasoning effort with automatic workflow orchestration, letting Claude decide per task whether to spin one up. Only on models that support xhigh effort.
• The bundled /deep-research workflow (requires the WebSearch tool).
• Ask Claude to write one directly; saved workflows become /<name> commands.

It requires Claude Code v2.1.154 or later and is available on all paid plans. It is on by default for Max and Team, opt-in on Pro via the Dynamic workflows row in /config, and off until an admin enables it on Enterprise. It runs across the CLI, desktop app, IDE extensions, headless mode, and the Agent SDK.

The demo that made people pay attention

The flagship example: Jarred Sumner used dynamic workflows to port Bun — roughly 750,000 lines of code — from Zig to Rust in 11 days from first commit to merge, keeping 99.8% of the existing test suite passing, with hundreds of agents running in parallel and two reviewers per file. Whether or not you care about Bun, that is the shape of the ideal use case: huge, parallelizable, and verifiable.

Where it is actually useful

• Codebase-wide bug sweeps and security audits, with independent agents checking each other.
• Large migrations — framework swaps or language ports across thousands of files.
• Cross-checked research, where claims get voted on before they reach you.
• Drafting a hard plan from several independent angles at once.

What it means for businesses building with AI

This is the same pattern we already use at Ascero AI: decompose a big job, run specialized agents in parallel, and put an adversarial reviewer in front of the output before it ships. Dynamic workflows makes that pattern native to Claude Code. For a small business, the practical upshot is not "run a thousand agents" — it is that verification is built into the loop. Built-in cross-checking is what makes automated work safe enough to put in front of a customer, which is the whole game.

The bottom line: the headline is hundreds of parallel agents, but the real advance is the adversarial verification and the script-as-orchestration model that keeps context clean on enormous jobs. Just do not believe the multi-day-unattended framing yet — the docs say otherwise.