Claude Code Dynamic Workflows: Orchestrate Hundreds of Agents via JavaScript

What It Is: Moving Planning from Conversation into Scripts

In standard Claude Code usage, the model keeps its plan in its head, deciding the next step each turn and funneling all sub-agent results back into the conversation context. Large tasks explode the context window and prevent true parallelism.

Dynamic Workflows change this: Claude writes a JavaScript orchestration script from a single prompt, hands it to an independent runtime that executes in the background, and schedules dozens to hundreds of sub-agents. Intermediate results stay in script variables, the parent session stays responsive, and only a single coordinated answer is returned at the end.

The official documentation highlights the key difference: sub-agents, skills, and agent teams are all directed by Claude turn-by-turn, whereas a workflow's baton is held by the script — so the orchestration can be saved and re-run.

Figure 1: Real-time progress view of a React→Solid migration workflow (source: Anthropic official blog)

Getting Started: Three Trigger Methods

ultracode keyword : Include "ultracode" in the prompt (prior to v2.1.160 the keyword was "workflow"), or simply say "use a workflow"; Claude Code will highlight and generate a script for that task.

/effort ultracode : A session-level switch that sets reasoning intensity to xhigh, letting Claude decide which tasks deserve a workflow — a single request may be split into coherent "understand code, implement, verify" workflows.

/deep-research : A built-in workflow that performs multi-angle retrieval, cross-verifies sources, and produces a cited report.

Prerequisite: Claude Code v2.1.154 or later. Pro users must enable manually in /config; Enterprise is admin-controlled; Max, Team, and API have it on by default. The first run shows the plan and raw script for confirmation.

Dissecting an Official Script

The documentation's audit example, saved as a workflow, looks like this:

export const meta = {
  name: 'audit-routes',
  description: 'Audit every route handler for missing auth checks'
}

const found = await agent('List every .ts file under src/routes/.', { schema: {...} })

const audits = await pipeline(found.files, file =
  agent(`Audit ${file} for missing authentication checks.`, { label: file }),
)

return audits.filter(Boolean)

meta

holds metadata; the body is plain JavaScript with top-level await: agent() spawns one sub-agent, pipeline() maps over a list spawning one per item, parallel() runs a group concurrently. agent() returns null on interruption, so the trailing filter(Boolean) removes empty slots.

Two important constraints:

Date.now() and Math.random() throw errors directly — this ensures re-runs reproduce the exact same call sequence.

Scripts cannot touch the filesystem or shell; all reads, writes, and executions are delegated to sub-agents.

More Reliable Results: Adversarial Verification Baked into Code

The value of workflows isn't just "more agents" — it's codifying quality patterns. In /deep-research, every conclusion must pass cross-voting by multiple independent agents; failures are filtered, unverifiable claims are marked unverified. In Bun's migration, each .rs file got two independent reviewer agents — letting agents rebut each other's findings, which is hard in a single pass but natural when encoded in a script .

Boundaries and Costs: What You Need to Know

Hard limits : max 16 concurrent agents, 1,000 total agents per run, 4,096 items per parallel / pipeline call.

No mid-run intervention : only permission prompts can pause; steps requiring human sign-off must be split into multiple workflows run sequentially.

Resume only within the same session : pausing works and completed agents read from cache; exiting Claude Code forces a full restart.

Token consumption is significantly higher : official guidance suggests a small trial run first; a "Large workflow" warning appears above 25 agents or an estimated 1.5 million tokens. /config lets you set scale tiers (default "medium" ≈ 15 agents), and lightweight phases can swap to a smaller model.

Satisfactory workflows can be saved as slash commands via /workflows and stored in the project's .claude/workflows/ directory for team-wide reuse.

Summary

Dynamic Workflows upgrade Claude Code's unit of work from "one conversation turn" to "one orchestrated run": plans become code, execution is observable, and results are reusable. Deploy them when a task exceeds a single context window or when the orchestration is worth repeating; for small fixes, regular chat and one or two sub-agents remain the most economical choice.

References

Official blog: https://claude.com/blog/introducing-dynamic-workflows-in-claude-code

Official docs: https://code.claude.com/docs/en/workflows