Claude Code Ultra Code and Dynamic Workflows

The Breakthrough

Claude Code introduced an 'Ultra Code' effort level that automatically triggers dynamic workflow orchestration, allowing the model to spawn specialized sub-agents and build custom execution harnesses for complex tasks rather than relying on a single, static context window.

How Dynamic Workflows Function

Dynamic workflows address common LLM failures like context rot, agentic laziness, and self-preferential bias by isolating tasks into separate, focused agent sessions.

• Automatic Orchestration: When set to /effort ultra, Claude Code evaluates the prompt and determines if a dynamic workflow is required, automatically selecting the appropriate pattern (e.g., Classify and Act, Fan Out and Synthesize, or Adversarial Verification).
• Custom Harnesses: The system generates a task-specific process at runtime. For example, a migration task might trigger sub-agents to read billing code, check provider documentation, calculate transaction volume, and run a 'devil's advocate' agent to verify the proposed solution.
• Adversarial Verification: Workflows incorporate dedicated verification agents that cross-reference findings against a rubric or codebase, significantly reducing the likelihood of false positives compared to standard single-session prompting.
• Manual Invocation: Users can force the behavior for specific tasks by typing /workflows before their prompt, ensuring the model uses a multi-agent approach even if it would have defaulted to a static harness.

Performance and Cost Considerations

Dynamic workflows are computationally expensive and token-heavy. In a deep research test involving 101 agents, the process consumed 3.7 million tokens over 11 minutes. While this approach provides higher accuracy for complex tasks like codebase-wide bug hunts or large-scale migrations, it requires monitoring via the /workflows command to track agent activity and token consumption in real time.