Members of the Claude Code team share firsthand: How to use dynamic workflows

Last week, Claude Code released a new capability: Dynamic Workflows.

This feature allows Claude to instantly write customized execution frameworks based on specific tasks, coordinate multiple sub - Agents to work in parallel, and solve the systematic failure problems in large - scale, highly parallel, and adversarial tasks.

Recently, Anthropic engineer Thariq published a long article sharing his initial workflow experiences and insights.

We have translated and organized the full text of his article.

Before delving into the technical details, Thariq provided some example prompts to help us understand the potential of workflows:

• “This test may fail once every 50 runs. Set up a workflow to repeatedly run the test, form hypotheses, and conduct adversarial verification on them in the work tree / Goal: Keep trying until a hypothesis succeeds.”
• “Use a workflow to review my last 50 sessions, identify recurring mistakes, and generate CLAUDE.md rules for these repetitive issues.”
• “Use a workflow to search the #incidents channel on Slack over the past six months and find the root causes of recurring issues that no one has submitted a ticket for.”
• “Take my business plan and run a workflow to have different Agents disassemble it from the perspectives of investors, customers, and competitors.”
• “There is a folder containing 80 resumes. Use a workflow to rank them, select the best candidates for the backend position, and review the top ten. Use the AskUserQuestion tool for interview scoring.”
• “I need to name this CLI tool. Use a workflow to generate multiple options and conduct a knockout tournament to select the top three best options.”
• “Use a workflow to rename our User model to Account.”
• “Review my blog post draft. Use a workflow to verify whether each technical statement complies with the codebase and ensure that no incorrect information is published.”

How Dynamic Workflows Work

Dynamic workflows execute a JavaScript file containing special functions that help generate and coordinate sub - Agents.

Meanwhile, dynamic workflows also include standard JavaScript features such as JSON, Math, and Array for data processing.

Dynamic workflows can determine the type of model an Agent uses and whether sub - Agents run in independent work trees, allowing Claude to select the required level of intelligence and isolation method.

If the workflow is interrupted, for example, by user operations or terminal exit, when resuming the session, the workflow can continue execution from the breakpoint.

Why Use Dynamic Workflows

When we use the default Claude Code framework to execute tasks, it needs to plan and execute simultaneously in the same context window. For many programming tasks, this is very effective, but in long - running, large - scale parallel, or highly structured adversarial tasks, problems may sometimes occur.

The reason is that the longer Claude processes complex tasks in a single context window, the more likely it is to experience the following types of failure modes:

• Agentic laziness: Claude may stop prematurely when handling complex multi - step tasks, claiming that the task is completed. For example, it may only process 20 out of 50 security reviews.
• Self - preferential bias: Claude tends to favor its own results or findings, especially when verification or evaluation is required.
• Goal drift: In multi - round operations, the task goal gradually deviates. Especially after compression and summarization, details such as edge cases or “do not do X” constraints may be lost.

Creating workflows can avoid these problems by assigning independent context windows to Claude instances for different goals, with each instance focusing on and isolating the task goal.

Differences between Dynamic and Static Workflows

You may have previously created static workflows using the Claude Agent SDK or claude - p to coordinate multiple Claude Code instances.

Static workflows need to account for all extreme cases, so they are usually more general. With the dynamic workflows of Claude Opus 4.8, Claude can now generate intelligent frameworks customized for your specific use cases.

Common Patterns of Dynamic Workflows

You can directly ask Claude to generate a dynamic workflow, or use the trigger word “ultracode” to ensure that Claude Code creates a workflow.

Understanding the common patterns of dynamic workflows helps to determine when to use them and how to guide Claude through prompts:

• Classify - and - act: Use a classifier Agent to determine the task type, then route to different Agents or actions based on the task. A classifier can also be used at the end to judge the output.
• Fan - out - and - synthesize: Split the task into multiple small steps, each handled by an Agent, and then summarize the results. This is especially suitable for a large number of small steps or when each step requires an independent context. The summarization step will wait for all distributed Agents to complete and then merge the structured outputs.
• Adversarial verification: The output of each sub - Agent is adversarially verified by another Agent against the evaluation criteria.
• Generate - and - filter: Generate multiple ideas, then filter them according to the evaluation criteria, remove duplicates, and only return high - quality and verified ideas.
• Tournament: Let multiple Agents execute the same task in different ways, and then compare the results pairwise through a judging Agent to select the best one.
• Loop until done: For tasks with unknown workloads, loop to generate Agents until the stop condition is met (no new discoveries or no more errors in the log), rather than a fixed number of rounds.

Use Cases

Migration and Refactoring

The rewrite of Bun from Zig to Rust was completed using workflows.

The key is to break the task into a series of small units that can be processed step by step, such as call points, failed tests, modules, etc. Each fix is assigned to a sub - Agent in an independent worktree; then another Agent conducts an adversarial review, and the changes are merged after confirmation.

If you want to run tasks in parallel as much as possible without overloading local resources, you can explicitly tell the Agent not to run commands that consume a lot of resources.

In - depth Research

We released an in - depth research skill (/deep - research) in Claude Code, which uses dynamic workflows.

Specifically, it conducts parallel web searches, fetches data sources, conducts adversarial verification of the statements, and finally integrates them into a research report with references.

However, this type of research is not limited to web searches. For example, you can also ask Claude to compile a status report from the Slack context or let it browse the codebase in - depth to study how a certain function is implemented.

In - depth Verification

On the other hand, if you have a report and want to verify each factual statement and its source one by one, you can build a workflow: first, an Agent is responsible for identifying all factual statements, and then a sub - Agent is derived to conduct a detailed verification of each statement. In addition, you can introduce a verification Agent to review the sub - Agent responsible for tracing the source to ensure that the sources it quotes are of high quality.

Sorting

You may have a batch of items that you want to sort according to a certain qualitative standard, which Claude Code is good at judging. For example, sort support tickets by the severity of bugs.

However, if you try to process more than 1000 lines in a single prompt, the quality is likely to decline, and the context may not be able to hold all the information. A better approach is to run a “tournament”: build a pipeline composed of pairwise comparison Agents. Pairwise comparison is usually more reliable than directly giving absolute scores.

You can also perform parallel bucket sorting first and then merge the results. Each comparison is completed by an independent Agent, and a deterministic loop is responsible for maintaining the entire tournament bracket. Only the current execution order remains in the context.

Memory and Rule - Following

If you find that Claude often misses or fails to execute a set of rules even after writing them in CLAUDE.md, you can create a dedicated workflow: list these rules and let a verification Agent check them one by one. Each rule corresponds to a verification Agent.

At the same time, create a sub - Agent with a skeptical perspective to review whether these rules are reasonable and whether they are truly aligned with the goals, which can reduce false alarms.

The opposite is also true: you can extract the problems that you repeatedly correct from recent sessions and code review comments; then let multiple Agents classify and organize them in parallel; then conduct adversarial verification on each candidate rule, for example, ask: Could this rule really have avoided a real error at that time? Finally, refine the verified rules back into CLAUDE.md.

Root Cause Investigation

The most effective way to debug is usually to propose several independent hypotheses first and then verify them one by one. However, if relying on only one context window, Claude is likely to fall into a certain “self - preference”: the more it looks, the more it believes its initial judgment.

Workflows can avoid this structurally. It can allow multiple Agents to propose hypotheses based on isolated evidence. For example, one Agent only looks at logs, one only looks at files, and one only looks at data. Then, each hypothesis is examined by a group of verifiers and refuters.

This method is not only applicable to code. It can also be used in sales scenarios, such as analyzing why sales declined in March; and in data engineering, such as investigating why a certain data pipeline failed. Any problem that requires a review and root cause analysis can be handled using a similar workflow.

Large - scale Ticket Sorting

Every team faces support ticket queues, bug reports, or other backlog tasks that often cannot be fully handled manually. A triage workflow can classify each to - do item, compare it with the tracked items to remove duplicates, and take corresponding actions. These actions may include trying to fix the problem directly or escalating it to a human user for processing.

In the triage workflow, “quarantine” is a very useful pattern. The core approach is to prohibit Agents responsible for reading untrusted public content from performing high - privilege operations; instead, these high - privilege operations will be performed by Agents specifically responsible for taking actions based on information. Combining the triage workflow with the /loop command allows Claude to continuously and automatically execute such tasks.

Exploration and Taste Judgment

Workflows are particularly useful when exploring different implementation paths for a solution, especially when the task involves subjective “taste” judgments (such as design or naming work) and needs to be evaluated based on a set of established criteria (Rubric).

Try asking Claude to explore and generate a series of potential solutions, then assign a “review Agent” and provide it with a clear set of evaluation criteria to define what a “good” solution is. When the review Agent determines that a solution fully meets the established criteria, the task is completed. In addition, solutions can be ranked or finally selected through a “tournament” - style competition mechanism based on this set of evaluation criteria.

Evaluation

You can run a lightweight evaluation process for a specific task: first, derive a group of Agents in an independent “worktree” to execute the task; then derive a group of “comparison Agents” to compare and score the specific output results generated by the previous Agents based on the established evaluation criteria. For example, you can use this mechanism to evaluate a Skill you created based on specific evaluation criteria and optimize it iteratively.

Model and Intelligent Routing

You can create a “classification Agent” specifically tuned for your task, which is responsible for deciding which base model to call to execute the task. This mechanism is particularly useful when your task involves a large number of tool calls. Through pre - analysis and research before formally executing the task, the classification Agent can accurately identify the most suitable base model for the current task.

For example, for the task of “explaining how the authentication module (Auth module) works”, the best choice of the base model is not fixed, but depends on the number of files in the authentication module and the overall structure of the entire codebase. At this time, the classification Agent can undertake the pre - analysis task and intelligently route the task to different base models such as Sonnet or Opus based on the judgment of the expected complexity of the task.

When Not to Use Dynamic Workflows

“Workflows” are a relatively new feature. Although in many application scenarios, they can bring significant results with less effort, not every task needs to rely on workflows; abusing workflows may lead to consuming far more Token resources than expected.