Claude Code Dynamic Workflows GA: Up to 1,000 Parallel Subagents with Ultracode

Anthropic has moved Claude Code Dynamic Workflows from research preview to general availability, expanding access to Pro plan subscribers for the first time. The feature lets Claude write its own orchestration scripts and coordinate up to 1,000 parallel subagents in a single run, with a hard cap of 16 agents running concurrently. For AI builders, this changes what is architecturally possible for autonomous code migrations, audits, and large-scale refactoring tasks.

What happened

Dynamic Workflows externalizes the orchestration plan into a JavaScript script executed by a background runtime. In previous Claude Code patterns, the model held the orchestration plan in its conversation context window, meaning every intermediate result accumulated until the window filled. Dynamic Workflows moves that logic out of the model's memory: each subagent receives a clean, focused context window, and the orchestration script handles branching, intermediate results, and verification loops. The model's context receives only the final synthesized answer.

The feature is available across all paid Claude Code plans: Pro ($20/month, off by default, enable through /config), Max and Team (on by default), and Enterprise (off by default, admin must enable). It runs on the Claude Code CLI, Desktop app, VS Code extension, and through the Claude API and cloud platforms including Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Why AI builders should care

The key architectural difference from simpler fan-out approaches is the adversarial verification layer. When Claude breaks a task into parallel subtasks and dispatches worker agents, a separate layer of adversarial agents challenges those findings before anything reaches the user. A migration agent that reports 99% completion is not accepted; a refutation agent is tasked with finding what it missed. The run iterates until findings converge under that adversarial check. This addresses a well-documented failure mode: models asked to verify their own work consistently over-report success.

For teams building AI-powered code migration or audit workflows, this means Dynamic Workflows is better suited for tasks that require agents to share findings and iterate toward a converged answer. OpenAI Codex, by contrast, runs parallel tasks in isolated cloud containers and returns pull requests without an explicit adversarial-review layer.

Practical implications

Token costs scale linearly with agent count. At Claude Opus 4.8 pricing of $5 per million input tokens and $25 per million output tokens, a full 24-hour parallel run can cost between $400 and $600. Anthropic warns that Dynamic Workflows "consume substantially more tokens than a typical Claude Code session" and recommends starting with a scoped task before attempting a full-codebase operation.

The Ultracode setting is a session-level effort mode that sets Claude to maximum reasoning and allows it to automatically decide when a task warrants a Dynamic Workflow. Anthropic recommends enabling it only for sessions where most tasks are large, ambiguous, or high-stakes. For routine work, drop back to standard effort to avoid running expensive workflows on tasks a single agent could handle faster and cheaper.

Caveats

The most widely cited proof point for Dynamic Workflows is the port of the Bun JavaScript runtime from Zig to Rust. However, Bun's creator Jarred Sumner has been an Anthropic employee since Anthropic acquired Bun in December 2025. The port is an internal Anthropic team demonstrating Anthropic's tool on Anthropic's own codebase - a genuine engineering achievement, but not independent third-party validation. The resulting Rust branch contains approximately 13,000 to 14,000 unsafe blocks, reflecting a line-by-line machine translation rather than an idiomatic rewrite. Full Mac and Windows test suite results are still pending.

Organizations adopting Dynamic Workflows should also plan for verification debt. As Mitch Ashley of The Futurum Group noted, "When one session fans out into hundreds of subagents across a migration or an audit, change outpaces what a team can review by hand. Organizations adopting this need verification, governance, and evidence capture that scale at the pace of generation."

FAQs

What is Claude Code Dynamic Workflows and how does it work?

Claude Code Dynamic Workflows externalizes the orchestration plan into a JavaScript script executed by a background runtime. The model coordinates up to 1,000 subagents per run, with a hard cap of 16 concurrent agents. Each subagent receives a clean context window, and the orchestration script handles branching, intermediate results, and verification loops. The model's context receives only the final synthesized answer.

How many parallel agents can Claude Code Dynamic Workflows spawn?

Official guidance indicates up to 1,000 total agents per run, with up to 16 agents running concurrently at once. These are hard caps enforced by the runtime.

Which Claude plans have access to Dynamic Workflows?

Dynamic Workflows is available across all paid Claude Code plans: Pro ($20/month, off by default, enable through /config), Max and Team (on by default), and Enterprise (off by default, admin must enable). Claude Code version 2.1.154 or later is required.

What are Ultracode and its recommended use cases in large sessions?

Ultracode is a session-level effort setting that sets Claude to maximum reasoning and allows it to automatically decide when a task warrants a Dynamic Workflow. Anthropic recommends enabling it only for sessions where most tasks are large, ambiguous, or high-stakes. For routine work, drop back to standard effort to avoid unnecessary resource use.