Claude Code Dynamic Workflows — Interactive Guide

💡 What Is a Dynamic Workflow?

Claude Opus 4.8 feature. Claude writes a JavaScript script that orchestrates hundreds of agents working in parallel. Results are merged and sent back to the main Claude session.

🎯 You describe the goal

→

📝 Claude writes JS script

→

🤖 N agents run in parallel

→

🔄 Results merged & returned

Unlike /goal (loop until done=true), workflows are width-first — many agents run horizontally, then stop. No endless loop checking.

🔑 The Core Insight

Skill = the HOW

A reusable recipe. "How do I do this task?"

Workflow = the HOW MANY / WIDTH

A parallel execution engine. "How many agents do I spin up?"

📊 /goal vs Workflow

✅ /goal — Depth-first

Loop until done == true
Can run 24h+
Good for: iterative refinement
Single agent or sub-agents
Checks end criteria each pass

⚡ Workflow — Width-first

N agents execute in parallel
Stops when all done
Good for: parallel batch jobs
Results synthesized at end
No criteria loop — "do your task, then stop"

Combine both: run a workflow inside a /goal for complex nested parallelism. But be careful — that's a money burner!

🧩 What Gets Generated

A JavaScript file saved to .claude/workflows/. You can:

View the raw script via /workflows command
Re-run it later without rebuilding from scratch
Browse all open workflows and stop them if needed

By default workflows may save to a global location, not your project. Specify the save path explicitly in your prompt.

📐 Feature Ladder — Complexity vs Power

As you move up: more complexity, more functionality, but also more cost and more risk.

Direct Question — Just ask

Quick questions, general knowledge work

Skill — Reusable Recipe

Repeatable tasks, define once, run many times

$$$

Sub-agent — Parallel Isolated Worker

Side tasks, isolated context, only reports back to main session

$$$

$$$$

Agent Team — Small Crew that Talks

War rooms, councils, debates. Members share task list & communicate

$$$$

/goal — Loop Until Done

Repeated passes until end criteria met. Can run 24h+

$$$$

$$$$$

Dynamic Workflow — Giant Parallel Job

Hundreds of agents, width-first, results synthesized at end

$$$$$

🔄 Sub-agent vs Workflow

Sub-agent

Single agent in parallel execution
Only talks back to main session
Cannot talk to other sub-agents
Isolated context window
Best for: one-off side tasks

Workflow

Many agents in parallel (can be 100+)
All results merged at end
More like a "big sub-agent"
Claude holds overall plan
Best for: batch parallel analysis

🏢 Agent Team vs Workflow

Agent Team

Members can talk to each other
Share task list
War rooms / councils / debates
Individual roles & tools
High cost due to inter-agent communication

Workflow

Agents work alone
No inter-agent communication
Results merged at end
No shared task list
More predictable cost

🆚 Feature Comparison Table

Feature	Agents	Inter-Agent Talk	Context	Best For
Direct Ask	1	—	Full session	Quick Q&A
Skill	1	—	Session + skill	Repeatable recipes
Sub-agent	1+	❌ No	Isolated	Side tasks
Agent Team	2-5	✅ Yes	Own + shared	Debates, councils
/goal	1 (+ subs)	Via subs	Session	Iterative until done
Dynamic Workflow	1 to 100+	❌ No	Own each	Batch parallel jobs

🔬 Deep Comparison: /goal vs Workflow

/goal — Depth Loop

Execution model: Sequential passes
End condition: Explicit done == true
Duration: Can run 24+ hours
Agent count: 1 primary (can spin up sub-agents)
Communication: Main agent coordinates sub-agents
Cost profile: Predictable but can balloon with long loops

Workflow — Width Parallel

Execution model: Parallel batch
End condition: All agents report done
Duration: Bounded by longest single task
Agent count: Up to 100+ simultaneously
Communication: None between agents
Cost profile: High input tokens (each agent reads own context)

🔍 Workflow Internal Architecture

👤 User prompt

→

📝 JS script created

→

🤖 Agent 1

🤖 Agent 2

🤖 ...N

🤖 Agent 1

🤖 Agent 2

🤖 ...N

→

🔄 Results merged

→

👤 Back to main session

Each agent is a full Claude call with its own context window. This is why input tokens add up fast — even with Haiku models.

💰 Real Cost: Skill Audit Case

Actual Data Nate Herk's Skill Audit

Haiku agents

Opus synthesis

~5M

input tokens

30+

minutes

~$100

half of $200/mo

One workflow prompt burned through half of a $200/month subscription. Budget carefully!

💸 Cost Drivers

Driver	Why It Costs	Mitigation
Number of agents	Each agent = 1 full Claude API call	Only use workflow when truly parallelizable
Context size per agent	Each agent reads its own context window	Minimize shared context; keep agent prompts lean
Model tier	Opus >> Sonnet >> Haiku in cost	Use Haiku for workers, Opus only for synthesis
Duration	Agents run concurrently but time still adds up	Bound scope tightly, name deliverables explicitly
Input >> Output	Most tokens are input (context reads)	Output-heavy is more expensive per token

✅ Cost-Saving Strategies

Haiku for all workers — Only the synthesis agent needs Opus
Bound the scope — Vague requests cause long loops
Name the deliverable — "Give me 3 specific fixes" not "improve my skills"
Check before running — Claude asks permission; review the plan first
View the raw script — See exactly what will run before approving
Stop runaway workflows — Use /workflows to see and stop running ones

📊 Model Cost Tiers (relative)

Model	Relative Cost	Use Case in Workflow
Haiku	$	Worker agents (scoring, analysis)
Sonnet	$$	Intermediate synthesis
Opus	$$$	Final synthesis, high-reasoning tasks
Ultra Code	MAX	Every prompt = workflow (very expensive!)

✅ Good Fit for Workflows

✅ Reviewing every file in a codebase
✅ 400+ file migration with independent subtasks
✅ High-risk work needing max compute/reasoning per piece
✅ Skill audit — 41 skills scored in parallel
✅ Feature ranking across multiple dimensions simultaneously
✅ Batch parallel research (especially with /deep research)

If the task can be split into N independent pieces that can run simultaneously, workflow is a fit.

❌ Bad Fit for Workflows

❌ Single edits or quick changes
❌ Quick questions or general Q&A
❌ Tasks that need sequential reasoning (use /goal instead)
❌ Work that doesn't parallelize well
❌ Everyday coding tasks (overkill)

If a simple skill or sub-agent would do the job, adding a workflow just burns money.

🔧 /deep research — Built-in Workflow

Slash deep research is a built-in workflow that:

Spins up multiple agents for parallel research
Agents vote on each claim
Outputs a cited deep research report

If you're doing research and want to build your own deep research skills, test /deep first to see how the built-in workflow works.

🚀 Ultra Code Mode

Uses X-high effort reasoning
Every single prompt runs as a workflow by default
Bypasses many permission checks
Most expensive mode — essentially X-high + workflow always on

Unless you need this level of orchestration constantly, use lower effort levels to control costs.

📋 The Skill Audit — Step by Step

Nate's actual workflow invocation prompt:

"Run a workflow that audits all of my Claude Code skills in two places: my user-level global skills and this project skills. Grade each on: Clarity, Front matter pass-fail, Trigger quality, and give me the single highest value fix. One final agent ranks every skill worst to best..."

⚙️ What Happened Internally

Claude asks permission
"Can you let me run a dynamic workflow to do this?"

User approves
Click "Yes, run it" — or view raw JS script first

41 Haiku agents spin up in parallel
Each scores one skill. All running simultaneously via JS script.

Opus synthesis agent receives all results
Ranks skills worst to best, generates HTML report with fixes

Output: HTML file saved to .claude/workflows/
Reusable — run again anytime by pointing at the workflow

👀 What You Can Monitor

🤖

Agent count

All 41 visible

🧠

Model used

Haiku per worker

📊

Tokens used

Per agent shown

⏱️

Duration

Running time

🛠️

Tools used

Per agent

Workflows may save to global ~/.claude/workflows/ by default. Specify your project path explicitly if you want project-level storage.

📁 Output Format

Nate's skill audit output was an HTML file containing:

All skills ranked worst to best
Specific fix recommendations per skill
Patterns identified across skills
Common pitfalls to avoid in future skill authoring

To re-run: /workflows → select the saved workflow → "run again". No need to rebuild the prompt.

🌳 Which Tool Should I Use?

Do you need immediate answers? → Just ask Claude

Is this a repeatable task you run often? → Use a Skill

Is it a messy one-off side task? → Use a Sub-agent

Do you need multiple agents to debate or collaborate? → Use Agent Team

Should it keep looping until a specific criteria is met? → Use /goal

Can the task be split into N independent parallel pieces? → Use Dynamic Workflow

Skills = HOW · Sub-agents = CLEAN SIDETASKS · Agent Teams = CREWS · /goal = LOOPS · Workflows = PARALLEL JOBS

🎯 The Parallelism Test

Ask yourself:

"Does this break into many pieces that can run individually of each other at the same time?"

If YES → Dynamic Workflow is appropriate
If NO → Consider /goal, skill, or sub-agent

⚡ Invocation Tips

Be explicit — "Set me up a dynamic workflow to do this" (not just "workflow" in natural speech)
The word "workflow" in natural speech just highlights in rainbow — doesn't auto-trigger
Claude will ask permission before running — say "Yes, run it"
You can always view the raw JS script before approving
Use /workflows to browse history and stop running ones

Ultra Code mode will bypass many permission checks automatically. Be extra careful about cost.

🔗 Combination Play

Advanced users can nest these:

/goal (loop)

→

Workflow (parallel)

→

Haiku agents

→

Opus synthesis

This level of autonomy is very powerful but very expensive. Use only when the value justifies the cost.

🔍

Dynamic Workflow

Claude Code feature (Opus 4.8+). Claude writes a JavaScript script that orchestrates many agents in parallel, merging results at the end. Width-first, not loop-based.

Skill

A reusable recipe for a task. Can be run by main session, sub-agents, agent teams, or workflows. Defines the "how."

Sub-agent

A parallel agent isolated from the main session context. Cannot talk to other sub-agents — only reports back to main session.

Agent Team

A small crew of agents that can talk to each other, share a task list, and have individual roles/tools. Good for debates, councils, collaborative problem-solving.

/goal

A Claude Code command that loops until done==true. Depth-first, iterative. Can run 24+ hours. Can spin up sub-agents. Good for: keep refining until criteria met.

Slash Deep Research (/deep)

Built-in workflow that spins up parallel research agents, has them vote on claims, and outputs a cited deep research report.

Ultra Code

Highest intelligence mode. Uses X-high reasoning + defaults every prompt to a workflow. Most expensive mode. Bypasses many permission checks.

Workflow Script

JavaScript file generated by Claude when a workflow runs. Saved to .claude/workflows/. Can be reviewed, reused, and re-run later.

Context Window

Each agent (sub-agent or workflow worker) has its own isolated context window. This is why input tokens multiply — every agent re-reads shared data.

Synthesis Agent

The final agent in a workflow that collects all parallel agent results and produces the unified output. Often Opus-tier for complex synthesis.

Width-first vs Depth-first

Width-first = many agents in parallel, results merged at end (workflow). Depth-first = one agent looping until done=true (/goal).

Cost-per-Agent

Each agent in a workflow = 1 full Claude API call with its own context. 100 agents = 100x the context reads = massive input token accumulation.