Dynamic Workflows: Orchestrating Parallel Agents

Dynamic Workflows allow Claude Code to manage complex tasks by offloading orchestration to a JavaScript-based runtime. Unlike standard subagent approaches that clutter the context window, these workflows move the plan into code, using script variables to store intermediate results. This keeps the main session responsive while the agent executes tasks in the background.

When a workflow is triggered (via the workflow keyword, ultracode setting, or /deep-research command), Claude dynamically decomposes the user prompt into subtasks. These are fanned out to up to 16 concurrent subagents. The system employs a refutation-based loop where agents independently verify findings until the results converge. The runtime enforces strict limits, capping runs at 1,000 subagents and preventing the workflow script from accessing the filesystem or shell directly—only the individual agents are granted those permissions.

Fast Mode: Optimizing for Speed

Fast Mode is a high-speed configuration for Claude Opus 4.8 designed for rapid iteration and debugging. It delivers 2.5x faster output token speeds while maintaining the model's intelligence. Priced significantly lower than standard mode (three times cheaper for Opus 4.8), it is intended for use cases where latency is prioritized over cost-sensitive autonomous tasks. Users can toggle this mode using the /fast command in Claude Code, indicated by a ↯ icon in the session.

Practical Application

Anthropic demonstrated the capability of Dynamic Workflows by porting the Bun runtime from Zig to Rust. The process involved hundreds of parallel agents mapping Rust lifetimes and generating behavior-identical code, resulting in 750,000 lines of Rust code and a 99.8% test suite pass rate over eleven days. This highlights the potential for using agentic workflows to handle massive, multi-file refactoring tasks that would be impractical to manage within a standard LLM context window.