← Back to feed

Claude Code can now write and execute its own orchestration scripts mid-task, spinning up tens to hundreds of parallel subagents within a single session. The feature is called Dynamic Workflows and ships today in research preview alongside Claude Opus 4.8.

What it actually does

Most coding agent failures are not model quality problems — they are scope problems. A single agent working serially cannot hold a 500,000-line codebase in context, run a comprehensive security audit, and verify every finding before surfacing results. Dynamic Workflows changes the architecture: Claude writes an orchestration script, deploys parallel subagents against different parts of the problem simultaneously, and runs independent verification on results before reporting back.

Three use cases Anthropic specifies:

1. Codebase-wide audits: Bug hunts, profiler-guided optimization, security hardening passes — Claude searches in parallel, then runs independent verification on every finding.
2. Large migrations: Framework swaps, API deprecations, language ports spanning thousands of files, executed end-to-end.
3. High-stakes verification: Claude runs independent attempts and deploys adversarial agents to break its own results before surfacing them. “When the cost of a wrong answer is high.”

The framing is direct: “Work you’d normally plan in quarters now finishes in days.”

Activation and availability

Dynamic Workflows are available in research preview across:

• Claude Code CLI, Desktop, and VS Code extension
• API (direct)
• Amazon Bedrock
• Google Vertex AI
• Microsoft Foundry

Access is limited to Max, Team, and Enterprise plan users. Enterprise deployment requires admin enablement.

Two activation paths:

• Explicit: Ask Claude to “create a workflow”
• ultracode mode: A new Claude Code setting accessible via the effort menu that sets effort to xhigh and lets Claude decide autonomously when a workflow is warranted

Anthropic’s caveat is notable: Dynamic Workflows “can consume substantially more tokens than a typical Claude Code session.” The recommendation is to start on a scoped task to calibrate usage. At $15/$75 per million tokens, a workflow that runs 100 subagents on a large codebase will generate a meaningful bill. This is an enterprise feature priced accordingly.

The architectural implication

Dynamic Workflows is a form of self-spawning multi-agent orchestration — the model writes its own scaffolding code, not just consults a fixed pipeline. This is architecturally distinct from tools like CrewAI or LangGraph, where the developer defines the agent graph. Here, Claude defines the graph based on the problem.

The practical limit is verification quality. Parallel subagents introduce consistency and coordination risks that serial execution does not have. Anthropic’s claim that Opus 4.8 “builds confidence before making big changes” is directly relevant here — the verifier agents are only as good as the model running them.

Whether Dynamic Workflows consistently outperforms well-scaffolded single-agent approaches (the conclusion of several recent agent architecture papers) remains to be tested at scale. The research preview designation suggests Anthropic is watching the same question.

Key Numbers

• Subagent scale: tens to hundreds per session
• Availability: Max, Team, Enterprise plans
• Activation: ultracode mode or explicit “create a workflow”
• Token cost: substantially higher than standard Claude Code sessions
• Deployment targets: CLI, Desktop, VS Code, API, Bedrock, Vertex AI, Microsoft Foundry