How to Use Claude Code’s Dynamic Workflows: Practical Tips from the Team

What Is a Dynamic Workflow?

Claude Code can now generate a dynamic workflow – a JavaScript file containing special functions that create and coordinate sub‑agents. These functions handle data with standard JavaScript utilities (JSON, Math, Array) and decide which model each agent uses, as well as whether agents run in isolated worktrees.

Why Use Dynamic Workflows?

When Claude runs complex tasks in a single context window, it can suffer from three failure modes:

Agentic laziness : the model stops early, claiming completion (e.g., processing only 20 of 50 security checks).

Self‑preferential bias : the model favors its own findings during verification.

Goal drift : after multiple summarisation steps, details and constraints may be lost.

Creating a workflow isolates each goal in its own Claude instance, keeping each sub‑task focused and preventing these issues.

Static vs. Dynamic Workflows

Static workflows (built with Claude Agent SDK or claude -p) must handle all edge cases up front and are therefore more generic. Dynamic workflows, powered by Claude Opus 4.8, generate a custom execution framework tailored to the specific use case.

Common Dynamic Workflow Patterns

Classify‑and‑act : a classifier agent routes tasks to specialized agents.

Fan‑out‑and‑synthesize : split a task into many steps, run agents in parallel, then merge structured results.

Adversarial verification : each sub‑agent’s output is checked by another agent against a verification rubric.

Generate‑and‑filter : produce multiple ideas, de‑duplicate, and keep only high‑quality, verified ones.

Tournament : run several agents on the same task and let a judging agent pick the best.

Loop‑until‑done : keep spawning agents until a stop condition (no new findings or errors) is met.

Illustrative Use Cases

Examples include:

Running a test that fails 1 in 50 times, looping until a hypothesis succeeds.

Analyzing the last 50 Claude sessions to extract recurring mistakes and generate CLAUDE.md rules.

Scanning a six‑month Slack #incidents channel to surface unseen root causes.

Ranking 80 resumes for a backend position and using an AskUserQuestion tool for interview scoring.

Renaming a User model to Account via a workflow.

Validating blog‑post claims against the codebase before publishing.

Deep Research Skill

The /deep‑research skill uses dynamic workflows to parallelise web searches, fetch sources, perform adversarial verification, and compile a cited research report. The same pattern can be applied to Slack context extraction or code‑base exploration.

Root‑Cause Investigation

Dynamic workflows enable independent agents to form hypotheses from logs, files, or data, then pass each hypothesis to a set of validators and challengers. This structure mitigates self‑bias and is useful for debugging, sales analysis, or data‑pipeline failure triage.

Large‑Scale Triage

A triage workflow classifies each ticket, de‑duplicates against existing items, and either attempts an automated fix or escalates to a human. The “quarantine” pattern isolates agents that read untrusted content from those that perform privileged actions, and can be combined with the /loop command for continuous operation.

Exploration & Judgment

When evaluating subjective solutions (e.g., design or naming), generate many candidates, then assign a reviewer agent a rubric to decide which meet quality standards. A tournament can rank or filter the final set.

Evaluation

Run a lightweight evaluation workflow by spawning agents in a separate worktree, then create comparison agents that score outputs against predefined criteria. This can be used to iterate on a custom skill.

Model Routing

A classification agent can decide which base model (e.g., Sonnet vs. Opus) to invoke based on task complexity, such as the number of files in an authentication module.

When Not to Use Dynamic Workflows

Dynamic workflows are powerful but can consume excessive tokens if overused. Simple coding tasks often do not need a multi‑agent pipeline.

Tips for Building Dynamic Workflows

Prompt design : detailed prompts that follow the techniques described yield the best results.

Combine /goal and /loop to set clear completion criteria and periodic execution.

Specify a token budget directly in the prompt (e.g., “use 10k tokens”).

Save workflows with the s key; they are stored under ~/.claude/workflows and can be packaged as a Skill for sharing.

When sharing as a Skill, place the JavaScript workflow files in the Skill folder and reference them in SKILL.MD. Treat the workflow as a template rather than a rigid script.

Dynamic workflows extend Claude Code’s capabilities and open many avenues for experimentation.