AB
AiBoss
project

Dynamic Workflows - A dynamic workflow technology introduced by Claude Code

Dynamic Workflows, introduced by Claude Code, is currently in research preview. It supports dynamically writing orchestration scripts in Claude, allowing for the parallel scheduling of dozens to hundreds of sub-agents within a single session...

What is Dynamic Workflows?

Dynamic Workflows, introduced by Claude Code, is currently in the research preview stage. It supports Claude in dynamically writing orchestration scripts, scheduling dozens to hundreds of sub-agents to work in parallel in a single session, automatically splitting tasks, verifying results, and resuming from breakpoints, compressing large engineering tasks that originally took weeks into a few days.

Main functions of Dynamic Workflows

  • Automatic task splitting and parallel executionClaude automatically breaks down large tasks into multiple subtasks based on user instructions and distributes them to parallel sub-Agents for simultaneous processing.
  • Results Verification and Counter-CensorshipThe results of each sub-agent are independently verified before being aggregated, and adversarial agents are deployed to actively challenge existing conclusions to ensure convergence of answers.
  • Resuming from a breakpointThe workflow automatically saves progress during operation, and can resume from the breakpoint if interrupted, without having to start from the beginning.
  • Full-process visualizationIt displays in real time the model used by each sub-Agent, token consumption, number of tool calls, and execution time.

The technical principles of Dynamic Workflows

  • Dynamic orchestration script generationClaude generates coordination scripts in real time based on task characteristics, rather than relying on predefined templates, enabling flexible task decomposition and scheduling.
  • Multi-level verification mechanismThe system employs a generation-review-iteration closed loop, where multiple agents independently attempt to solve the problem from different perspectives, while other agents are responsible for refuting and verifying the results until convergence is achieved.
  • External coordination architectureThe coordination logic operates outside the dialogue, ensuring that the plan can proceed steadily even as the task scales up, unrestricted by the dialogue context.
  • Long-term operation supportThe design supports continuous operation for hours to days, adapting to the time requirements of large-scale engineering tasks.

How to use Dynamic Workflows

  • Direct command startIn Claude Code, you can directly request Claude to create a workflow, for example, by entering "Create a workflow" to trigger automatic orchestration.
  • Enable Ultracode modeEnable via the Effort menu ultracode By setting the effort to xhigh, Claude will automatically determine when to trigger the workflow to process complex tasks.
  • Platform and permissions:
    • Compatible with Claude Code CLI, desktop, VS Code plugin, as well as Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.
    • Max, Team, and API users have this feature enabled by default; Enterprise users need to have it manually enabled by the administrator in the settings.

The core advantages of Dynamic Workflows

  • Scalable processing capabilitiesA single session can coordinate the parallel work of hundreds of agents, making it suitable for large-scale codebase-level operations across files and services.
  • High reliability outputMultiple verification and adversarial review mechanisms significantly reduce error rates, making them suitable for critical tasks in high-cost error scenarios.
  • End-to-end automationThe entire process, from task planning to implementation and verification, is completed automatically without the need for manual intervention in orchestration or management of sub-agents.
  • Real-world verification casesJarred Sumner used Dynamic Workflows to port a Bundle from Zig to Rust, generating approximately 750,000 lines of code with a 99.8% pass rate in the test suite, taking only 11 days from the first commit to the merge.

The project address for Dynamic Workflows

  • Project official websitehttps://claude.com/blog/introducing-dynamic-workflows-in-claude-code

Comparison of Dynamic Workflows with similar competitors

Dimension Claude Code Dynamic Workflows OpenAI Codex CLI
Parallel Scale Hundreds of sub-agents operate in massive parallelism, making them suitable for codebase-level refactoring. Sub-agents have limited parallel scale and are more suitable for module-level tasks.
Task orchestration Dynamically and automatically generate orchestration scripts without manual intervention. It requires manual triggering or configuration by the user, resulting in a low degree of orchestration automation.
Verification mechanism Built-in adversarial review and multi-round iterative verification ensure high reliability. Relying on auto-review and user approval, the verification process is more traditional
runtime Natively supports running for several days or more, with breakpoint resume capability. Codex Cloud supports long-running background processes, but local sessions are limited.
Token efficiency High overhead (multiple agents operating in parallel have high overhead) Token consumption is low (approximately 1/3 to 1/4 of Claude's).
Terminal performance Leading in complex multi-file tasks such as SWE-bench Pro Leading in terminal programming tasks such as Terminal-Bench 2.0
Usage Mode Emphasizing "cyber contractor" style collaboration, developers can see the entire process. The emphasis is on "fire-and-forget" unattended operation, with the process completed automatically in the background.

Application scenarios of Dynamic Workflows

  • Full codebase vulnerability scanParallel search of the entire service or repository, run independent verification on each discovery, and output a real issue report.
  • Large-scale code migrationHandling end-to-end migrations involving thousands of files, such as framework replacement, API deprecation, and language porting.
  • Dual verification for critical missionsBefore making high-risk decisions, ensure the reliability of the conclusions through independent trials and adversarial challenges.
  • Performance optimization auditBased on the analyzer's guidance, it reviews dead code, performance bottlenecks, and cleanup opportunities in the codebase in parallel.