Claude Code ultracode — What It Is, How to Enable It, and Who Can Use It
Claude Code's ultracode — What It Is, How to Enable It, and Who Can Use It
ultracode, which surfaced in late May 2026, has become one of the most talked-about additions among Claude Code users. Yet the information out there is fragmented — questions range from "is this a slash command or an inference level?" to "how many tokens does it burn?" to "can I use it on my plan?" This post consolidates everything: what ultracode actually is, how its mechanism works, exactly how to enable it, and what the cost and plan-level availability look like. The short answer: ultracode is not a standalone feature — it is an effort (inference-level) setting that automatically activates Claude Code's new orchestration engine, Dynamic Workflows, for the entire session.
🧩 The Relationship Between ultracode and Dynamic Workflows
Alongside the release of Claude Opus 4.8, Anthropic introduced Dynamic Workflows — an orchestration engine built into Claude Code — together with the ultracode effort setting that activates it automatically. The feature is currently in research preview, not yet generally available, and requires Claude Code v2.1.154 or later (run claude update to upgrade).
▶ Dynamic Workflows = the execution engine. Claude externalizes its orchestration plan as a JavaScript script, and the runtime drives tens to hundreds of agents in parallel according to that script.
▶ ultracode = the effort level that keeps that engine running automatically throughout the entire session. Internally it combines xhigh inference depth with automatic workflow orchestration.
Here is how the two concepts interlock at a glance.
graph LR
A[/effort ultracode
= xhigh effort] --> B[Dynamic Workflows
engine activated]
B --> C[JS orchestration
script generated]
C --> D[Parallel agent execution
up to 16 concurrent]
style A fill:#e8f8f5,stroke:#16a085,color:#117a65
style B fill:#fef9e7,stroke:#f39c12
style C fill:#eaf2f8,stroke:#2980b9
style D fill:#eafaf1,stroke:#27ae60,color:#1e8449
🔗 Diagram summary: Enabling ultracode activates xhigh inference, which starts the Dynamic Workflows engine. Claude generates a JS orchestration script, and the runtime executes it with up to 16 agents running concurrently.
The structural differences from traditional multi-agent approaches are summarized in the table below.
| Dimension | Traditional Subagents | Dynamic Workflows |
|---|---|---|
| Planning authority | Claude decides turn-by-turn | Script (code) drives decisions |
| Intermediate result storage | Claude's context window | Script variables |
| Scale | Small number of delegations | Tens to hundreds of agents |
| Re-execution | Not supported | Script is saved and re-runnable |
| Resume after interruption | Restart from the turn | Resume within the same session |
The core insight is "externalizing the plan." With traditional multi-agent setups, all results accumulate in Claude's context window and the model re-evaluates coordination logic on every turn. Dynamic Workflows moves that coordination logic into code, breaking through the single-context-window bottleneck. The analogy: the old approach is a director who keeps every blocking note in their head; the new approach is a director holding a script who can move dozens of actors simultaneously.
⚙️ How It Works — Runtime Behavior and Constraints
When a task is submitted, Claude writes a JavaScript orchestration script and the runtime executes it in the background while the user session remains responsive. The documented runtime constraints are as follows.
✓ Up to 16 agents running concurrently (may be lower depending on CPU core count)
✓ Hard cap of 1,000 total agents per run — a backstop against infinite loops
✓ The workflow script itself has no direct access to the filesystem or shell. All I/O must go through agents
✓ Completed agent results are cached — resuming after an interruption skips already-finished work
Progress is visible via the /workflows command or the summary row in the task panel below the input field. A running workflow can be interrupted at any time (completed results are preserved). The deliberate design choice to keep the script isolated from the filesystem and shell means that even with hundreds of agents running in parallel, permission boundaries remain well-defined — a meaningful safety property.
🚀 How to Enable It — Three Entry Points
The claim that "you enable it as an inference level, not a slash command" specifically refers to Method 2 below, and it is accurate. There are three entry points in total, and which path activates automatically depends on the nature of the task.
flowchart TD
A([Task request]) --> B{effort set to
ultracode?}
B -->|YES| C[Automatically assess complexity
and generate workflow]
B -->|NO| D{Does the prompt
contain "workflow"?}
D -->|YES| E[Generate a one-off
workflow]
D -->|NO| F[Process as a
standard response]
style A fill:#3498db,stroke:#2980b9,color:#ffffff
style B fill:#fef9e7,stroke:#f39c12
style D fill:#fef9e7,stroke:#f39c12
style C fill:#eafaf1,stroke:#27ae60,color:#1e8449
style E fill:#eafaf1,stroke:#27ae60,color:#1e8449
style F fill:#fdedec,stroke:#e74c3c,color:#c0392b
🔁 Diagram summary: On each task, if effort is ultracode, Claude automatically assesses complexity and generates a workflow. Otherwise, if the prompt contains the word "workflow," a one-off workflow is generated. If neither condition is met, the request is handled as a standard response.
Method 1 — Include the "workflow" keyword (manual, one-off)
Including the word workflow anywhere in a prompt triggers workflow script generation for that request only. The session effort level is not changed. If it triggers unintentionally, press alt+w to suppress it for that prompt.
Method 2 — /effort ultracode (automatic, session-wide) ★ Primary
Claude automatically decides whether to generate a workflow for every substantive task in the session. Simple tasks get standard responses; complex ones are automatically handled as workflows.
✓ Resets when the session ends — the next session starts at the default effort level
✓ For routine tasks, it is recommended to drop back with /effort high
✓ On models that do not support xhigh, the ultracode option simply does not appear in the /effort menu — it is exposed only on supported models such as Opus 4.8
Method 3 — /deep-research (specialized, research-only)
A built-in workflow that requires no additional configuration. It fans out web searches across multiple angles, cross-validates sources, and returns a cited report. This is the lightest entry point for getting parallel verification benefits on research tasks without enabling ultracode session-wide.
🎯 When Is It Actually Useful
Official documentation and multiple technical publications converge on three primary use cases.
🟢 Large-scale codebase audits — bug hunting, security vulnerability review, performance optimization. Agents independently review different areas and then adversarially verify each other's findings, increasing overall confidence.
🟢 Large-scale migrations — code transformations spanning hundreds to thousands of files. Work that exceeds any single context window is split and processed in parallel.
🟢 Multi-perspective cross-validation research and design — multiple agents independently produce plans that are then compared and synthesized. Useful for evaluating a new architecture from several independent viewpoints before committing.
🔴 Not a good fit: simple edits, short questions, single-file changes. Even the xhigh reasoning level alone consumes tens of thousands of tokens of thinking budget per request. Enabling ultracode for lightweight work is pure cost with no benefit.
💰 Token Cost and Plan-Level Availability
Availability and the activation path differ by plan. The following is based on official documentation.
| Plan | Monthly Price | Availability | Activation |
|---|---|---|---|
| Pro | $20 | Available | Must be manually enabled via /config |
| Max 5x | $100 | Enabled by default | No additional configuration needed |
| Max 20x | $200 | Enabled by default | No additional configuration needed |
| Team | $30 per user | Enabled by default | No additional configuration needed |
| Enterprise | Custom pricing | Disabled by default | Admin enables it in the admin settings panel |
Comparing the estimated 5-hour window token budgets across plans as a bar chart makes the differences concrete. This is useful for gauging the weight of a single workflow run.
※ The 5-hour window token figures are third-party estimates. On Pro, a single workflow run can consume a significant fraction of the budget. Max 20x has headroom even for large-scale tasks.
Official documentation does not publish specific token consumption figures — it describes usage only as "meaningfully more tokens than a single-agent approach." Third-party analysis estimates approximately 7x token consumption compared to a single-agent session (treat this as a rough estimate, not a hard number).
/model to check the active model and, for pipeline stages that do not require heavy reasoning, specify in your task description that they should be routed to a smaller model. Interrupting a running workflow via /workflows still preserves all completed agent results.
To disable workflows, use any of three methods: ① toggle it off in /config, ② add "disableWorkflows": true to ~/.claude/settings.json, or ③ set the environment variable CLAUDE_CODE_DISABLE_WORKFLOWS=1. Workflow usage counts against the same plan usage and rate limits as standard requests.
📌 Summary and Takeaways
▶ ultracode is not a standalone feature. It is an xhigh effort level that automatically activates the Dynamic Workflows engine for the entire session. "Enable it as an inference level" is accurate, and the exact command is /effort ultracode.
▶ The core value proposition is externalized planning combined with parallel scale (up to 16 concurrent, up to 1,000 agents per run). The primary targets are codebase audits, migrations, and cross-validation research that exceed a single context window.
▶ Recommended adoption sequence: ① First test the effect with a one-off workflow keyword prompt → ② Once validated, enable /effort ultracode only for heavy sessions → ③ Drop back to /effort high when done.
On plan choice: Max 20x is the most practical fit given its token headroom; Max 5x is the minimum for viable use. Pro can enable it via /config, but the small window (~44k) is a significant constraint for anything at scale.
Caveats: The ~7x token multiplier and per-plan window token figures are third-party estimates, not official numbers. Plan conservatively when budgeting costs. As a research preview, constraints such as the agent cap and effort-level exposure conditions may change in future releases.
References
• Anthropic Official Blog — Introducing Dynamic Workflows in Claude Code (claude.com/blog/introducing-dynamic-workflows-in-claude-code)
• Claude Code Official Documentation — Workflows (code.claude.com/docs/en/workflows)
• ThePlanetTools — Opus 4.8 Ultracode Guide
• MarkTechPost — Opus 4.8 Release Coverage
Some figures in this post (token consumption multiplier, per-plan window limits) are third-party estimates and may not reflect actual values. The feature is in research preview and specifications are subject to change — consult the official documentation before adoption.
Collecting and organizing software development resources firsthand, with a second pass before publishing.
This post is based on publicly available data and cited sources. Last updated: June 8, 2026
댓글
댓글 쓰기