Claude Code Ultracode: Run Parallel Agents at Scale With One Setting

If you've used Claude Code's dynamic workflows since they launched on May 28, 2026, you may have noticed something: the keyword that used to trigger them stopped working. That's not a bug — Anthropic renamed the trigger from "workflow" to "ultracode" in v2.1.160, and at the same time turned it into a first-class /effort setting.

This guide explains exactly what Ultracode is, how it differs from both regular Claude Code and UltraPlan, when to use it, and how to keep costs from spiraling on large sessions.

What Is Claude Code Ultracode?

Ultracode is not just a higher effort level. It is a two-part setting that does two distinct jobs simultaneously:

The combination means you give Claude a complex, multi-step engineering goal and it handles the decomposition, parallelization, and verification automatically. A single prompt like "migrate our REST API to GraphQL, update all tests, and document the changes" can become three sequential workflows: one to map the codebase, one to make the changes in parallel across services, and one to verify tests pass.

Quick disambiguation: Ultracode is not the same as Ultraplan. Ultraplan moves the planning phase to a cloud editor where you review and approve before anything executes. Ultracode stays in your local terminal and executes autonomously. Use Ultraplan when you need human approval at each step. Use Ultracode when you trust Claude to run with a well-specified goal.

How Dynamic Workflows Work Under the Hood

When Ultracode is active, Claude doesn't just write code — it writes an orchestration script first. That script defines:

• Which subtasks can run in parallel vs. sequentially
• How many subagents to spin up (up to 16 concurrent at a time)
• How results are validated before moving to the next stage
• Where to checkpoint so the session can resume if interrupted

The runtime enforces two hard limits:

These caps are per-run, not per-session. A session can contain multiple workflow runs in sequence. So a task that maps your codebase (run 1), then patches 200 files in parallel (run 2), then runs integration tests (run 3) stays well within the 1,000-agent cap per run while doing thousands of operations in total across the session.

How to Activate Ultracode

There are two ways to turn on Ultracode in Claude Code v2.1.160 and later:

Method 1: Slash command (persists for the session)

/effort ultracode

Claude will confirm that Ultracode is set and that the effort level will persist for the rest of the session. Every substantive request in that session will automatically get workflow orchestration treatment.

Method 2: Keyword in your prompt (single task)

Include the word ultracode anywhere in your prompt:

Ultracode: refactor the payment module to use the new idempotency keys and 
update every integration test.

Claude recognizes the keyword and writes a workflow for that task only, without changing the session's default effort level.

Step-by-Step: Your First Ultracode Session

Here's a practical walkthrough for a real engineering task.

Step 1: Start with a scoped read-only pilot

/effort ultracode
Ultracode: map the Express 4 API surface we're using — list every 
deprecated method call, middleware pattern, and router config. 
Don't change anything yet.

Claude runs a read-only workflow across your codebase. Review the output before proceeding.

Step 2: Approve and proceed to the change phase

Once you've reviewed the map:

Good. Now make the changes. Prioritize router and middleware first, 
then deprecated response methods. Run tests after each batch.

Claude writes a new orchestration script, fans out parallel agents to handle different parts of the migration, and surfaces a summary when done. You'll see the workflows in the agents view (claude agents --json shows the current blocking state if anything stalls).

Step 3: Verify and downgrade effort

After the migration:

/effort high
Run the full test suite and summarize any remaining failures.

Drop back to high effort for verification tasks. You've spent xhigh tokens on the hard part — no need to keep burning them on a npm test interpretation.

Ultracode vs. Other Effort Levels

The key insight: xhigh effort without Ultracode makes Claude think harder on each individual response. ultracode makes Claude think harder and coordinate multiple agents. The difference in cost and time is significant — Ultracode is appropriate for tasks that genuinely span multiple files, services, or systems.

Cost Management: Keep Ultracode From Running Away

Each workflow run consumes substantially more tokens than a standard Claude Code request. Here's what to do:

Which Plans Support Ultracode?

Dynamic workflows — and therefore Ultracode — are available on every paid Claude Code plan as of June 2026:

• Pro — available
• Max (5x and 20x) — available, higher monthly limits
• Team — available
• Enterprise — available, with configurable concurrency settings

Free plan users do not have access to dynamic workflows. Note that Ultracode sessions consume credits from your standard Claude Code usage pool (not the new Agent SDK credit pool introduced June 15, 2026 — unless you're calling Claude programmatically via the Agent SDK).

What the Agents View Shows You

One of the quieter improvements in v2.1.162 pairs well with Ultracode: the claude agents --json flag now exposes what each waiting session is blocked on. During a complex workflow, subagents often wait for earlier tasks to finish before they can start. Previously this looked identical to an agent doing nothing. Now you see the dependency explicitly:

json{
  "session": "wf-0029",
  "status": "waiting",
  "blocked_on": "task:auth-schema-migration",
  "reason": "downstream agent waiting for schema lock release"
}

This is especially useful during long Ultracode sessions where tens of agents are running in the background. If a workflow is slower than expected, check --json before assuming something is wrong.

Key Takeaways

• Ultracode = xhigh effort + automatic workflow orchestration. It's a compound setting, not just a deeper reasoning mode.
• The keyword changed. "ultracode" now triggers dynamic workflows; "workflow" no longer works as of v2.1.160.
• Hard limits: 16 concurrent agents, 1,000 total per workflow run.
• Activate with /effort ultracode (session-wide) or by including "ultracode" in your prompt (single task).
• Cost control: pilot with read-only tasks first, scope tightly, downgrade effort after the hard work is done.
• Fast mode on Opus 4.8 is 3x cheaper than before — use it for large workflows to control costs.

Next Steps

If you're preparing for the Claude Certified Architect (CCA) exam, dynamic workflows and agentic coordination are core topics. Our CCA practice test bank includes 200+ questions on Claude Code's agentic features, effort controls, and orchestration patterns — with detailed answer explanations for every question.

Already comfortable with Ultracode? Check our guide on Claude Code GitHub Actions integration to run workflows in CI/CD pipelines, or Claude multi-agent orchestration patterns for building your own agent networks outside of Claude Code.