Claude Code ultracode is two features that share one name. One is a keyword you type to run a single task as a dynamic workflow. The other is a session toggle that tells Claude to plan a workflow for every substantive task. Neither one is a “thinking mode”, and neither one fires the way most people expect once Claude Code runs without a human at the keyboard.
That second part is my whole day job. I run 10+ brand containers where Claude Code works on cron, unattended, through claude -p. The post you’re reading was written by one of those scheduled runs at 00:05 on 20 Nov 2026. So this isn’t a laptop demo. It’s the docs (checked today against Claude Code v2.1.328) plus what my own container showed when I went looking.
Here’s what you’ll get: what ultracode actually does, how it differs from ultrathink, xhigh and max, how to turn it on and off, the headless table nobody hands you, what it costs, and the rule I use to decide whether it gets anywhere near a fleet.
What Claude Code Ultracode Actually Is

Start with the thing underneath it: a dynamic workflow. Claude writes a short JavaScript script that orchestrates many subagents at once. A runtime executes that script in the background while your session stays responsive. The script holds the loop, the branching and the intermediate results, so Claude’s context only holds the final answer. Anthropic shipped dynamic workflows in v2.1.154 on 28 May 2026.
Ultracode is how you ask for one. It comes in two forms:
- The keyword. Put
ultracodein a prompt you type, and Claude runs that one task as a workflow instead of working through it turn by turn. The session’s effort level doesn’t change. Typing “use a workflow” in plain words counts as the same opt-in. (The trigger word used to be “workflow”. It was renamed toultracodein v2.1.160 on 2 June 2026, which is why older tutorials disagree with each other.) - The toggle. Run
/effort ultracodeand Claude decides for itself, task by task, when to plan a workflow, for the rest of the session. One request can turn into several workflows in a row: one to understand the code, one to change it, one to verify it.
The docs put it plainly: ultracode is “a Claude Code setting rather than a model effort level.” It doesn’t make the model think harder. It changes who holds the plan. With subagents and skills, Claude decides turn by turn what happens next. With a workflow, the script decides.
Workflows run on all paid plans, on the Anthropic API, and on Bedrock, Google Cloud and Microsoft Foundry. On Pro, you first switch them on from the Dynamic workflows row in /config. Ultracode also needs a model that supports xhigh effort. That rules out Opus 4.6 and Sonnet 4.6. Opus 5.5, Sonnet 5.5, Haiku 5.5, the Fable models, Opus 5, Opus 4.8 and Opus 4.7 all qualify.
Ultracode vs Ultrathink vs xhigh vs max

This is where most explainers drift. At least one of the pages I read for this post frames ultracode as an “X-high effort” thinking mode. That was close to true before v2.1.284. It isn’t now. Four different switches, four different jobs:
| Switch | What it is | What it changes | How long it lasts |
|---|---|---|---|
ultrathink | A keyword | Runs that one turn at xhigh effort (interactive terminal, Opus 5.5 / Sonnet 5.5 / Haiku 5.5 / Fable 5.1 via Anthropic’s API, since v2.1.311) | One turn |
xhigh | An effort level | Deeper reasoning per step at higher token spend | Session or saved default |
max | An effort level | The deepest reasoning. The docs warn it “is prone to overthinking” | Current session only (unless set by env var) |
ultracode keyword | An opt-in | Runs one task as a dynamic workflow (many agents, one script) | One task |
/effort ultracode | A setting | Claude plans a workflow for every substantive task, at whatever effort level you’re on | Session, or every session via the ultracode setting |
The easy way to keep them apart: effort is how long one agent thinks. Ultracode is how many agents work. They stack. You can run ultracode at medium effort, and since v2.1.284 (28 Sep 2026) it stays on when you change the level. Before that, turning ultracode on forced xhigh, and picking another level silently turned it off.
One exception catches people out. Launching with claude --effort ultracode (v2.1.203+) or sending the Agent SDK value effortLevel: "ultracode" turns ultracode on and sets the level to xhigh. The other routes leave your level alone.
How to Turn Ultracode On, Dismiss It, and Turn It Off

Every route, from the docs as of 20 Nov 2026. Commands live in my Claude Code cheat sheet too. This is the ultracode slice.
| You want to… | Do this | Version note |
|---|---|---|
| Run one task as a workflow | ultracode: audit every route under src/routes/ for missing auth checks | Keyword since v2.1.160 |
| Cancel the keyword on this prompt | Option+W (macOS) or Alt+W, or backspace right after the highlighted word | |
| Stop the keyword from ever triggering | Turn off Ultracode keyword trigger in /config | |
| Ultracode for this session | /effort ultracode, or Tab in the /effort slider, then Enter | Toggle form v2.1.284+ |
| Start a session with it on | claude --effort ultracode (also sets xhigh) | v2.1.203+ |
| Every session | "ultracode": true in a settings file | Not via effortLevel or CLAUDE_CODE_EFFORT_LEVEL |
| Turn it off | /effort ultracode off | Before v2.1.284 this failed with Invalid argument |
| Kill workflows entirely | "disableWorkflows": true or CLAUDE_CODE_DISABLE_WORKFLOWS=1 | Also removes ultracode |
| Watch or stop a run | /workflows, then p to pause, x to stop, s to save the script as a command |
Two details worth knowing before you flip anything. First, effortLevel and CLAUDE_CODE_EFFORT_LEVEL don’t accept ultracode. People try that, nothing happens, and they conclude the feature is broken. Second, there’s no admin switch that blocks ultracode on its own. An org can disable workflows entirely, which removes ultracode with them. Short of that, anyone can turn it on wherever it’s available.

Get the AI Playbook
The skill files, permission settings and run logs behind my unattended Claude Code fleet, including the guardrails I use before anything fans out. Free, and you can leave whenever you like.
Ultracode in Headless and Cron Runs: What Fires and What Doesn’t

The better explainers out there (I read six for this post) get the keyword rule right: it doesn’t fire from -p. What I didn’t find in any of them is the operator’s version, meaning what that rule plus the setting, bypass mode and the usage limit add up to when nobody is watching. So here it is in full.
The keyword is an opt-in only in a prompt a human types: at the interactive prompt, in an IDE panel, in a Remote Control client, or in an Agent SDK app that stamps keyboard input as origin: { kind: "human" }. Since v2.1.210, webhook payloads and relayed PR comments no longer trigger it either. That’s a safety feature. A webhook or a PR comment shouldn’t be able to launch a hundred agents on your bill.
| Where the prompt comes from | Does the ultracode keyword fire? | Can a workflow still start? |
|---|---|---|
| You, typing interactively | Yes | Yes, after the approval prompt (it depends on permission mode) |
claude -p "…" (cron, CI, scripts) | No | Only if the permission check lets the Workflow tool through: an allow rule, auto mode, bypass mode, a PreToolUse hook, or --permission-prompt-tool |
| Agent SDK, input not stamped human | No | Same rules as -p, or a canUseTool callback |
| Scheduled task prompt | No | Yes if the session has the ultracode setting on (it’s announced to headless turns since v2.1.299) |
| Webhook payload or relayed PR comment | No | Only through the session’s own settings and permissions |
Here’s the part that should make fleet operators sit up. The keyword is inert headless, but the setting is not. If "ultracode": true sits in a settings file your cron sessions read, every scheduled run gets ultracode. Pair that with bypass permissions, which is how plenty of unattended setups run, mine included, and nothing stands between “Claude decided this task deserves a workflow” and “the workflow is running”. The docs’ approval table says it outright: in bypass mode, “Claude Code doesn’t prompt you. The run starts immediately.”
Then there’s the usage-limit behaviour, which is the real trap. Since v2.1.271, a workflow that hits your claude.ai usage limit pauses and resumes after the reset. But only in an interactive terminal session or a local Code-tab session, on a subscription without usage-based billing, with autoContinueAtUsageLimit on. In claude -p, the Agent SDK, background sessions and scheduled-task sessions, the agents that hit the limit fail instead. Inside a pipeline(), the run carries on with null where their results should be. Your skill then works from a partial answer and may still report success. That is exactly the failure class I spend my time catching in unattended orchestration.
What I Found When I Checked My Own Fleet

Plain disclosure first. I did not run an interactive ultracode session for this post, so there are no agent counts or token totals here. This post was written by a scheduled claude -p run, which is precisely the situation where the keyword can’t fire. I also tried to start a nested headless session to test the keyword end to end. It failed at authentication before the model ever ran. So everything below is what the container showed me, not a benchmark.
The dispatcher line
Every skill in my fleet starts the same way. Line 182 of cron-dispatcher.sh:

⚡ GET THE AI EDGE
Weekly AI tips that actually save you time and money. No fluff, no hype — just what works.
claude --print --dangerously-skip-permissions -p "$INSTRUCTION" < /dev/null
That’s -p, so the keyword is dead on arrival. It’s also bypass mode, so a workflow could start if Claude chose to write one. Nothing in my settings asks for that. ~/.claude/settings.json holds two keys (skipDangerousModePermissionPrompt and model), with no ultracode, no Workflow allow rule and no disableWorkflows. There’s no .claude/workflows/ folder anywhere, and ultracode appears in 0 of my 39 skill files.
What the CLI reported at startup
- Version: the container runs v2.1.292. The changelog is at v2.1.328 (19 Nov 2026). That’s 36 releases behind, and the headless-announcement fix (v2.1.299) is one of the releases it doesn’t have yet.
- The Workflow tool is loaded in
-p. The init event of a headless session listsWorkflowamong its tools, withdeep-researchandworkflow-authoringamong the slash commands. Workflows aren’t an interactive-only feature, and the docs confirm they run in-pand the SDK. --effort ultracodeis accepted.claude --helpon v2.1.292 lists onlylow, medium, high, xhigh, max. Pass--effort bogusand it warns “Unknown –effort value… Valid values: low, medium, high, xhigh, max.” Pass--effort ultracodeand it says nothing. The help text undersells the binary.- This run’s own session: effort
medium, unattended, 8 CPUs in the container. The 16-concurrent-agent default scales down on CPU-limited containers, so the real ceiling here may be lower.
The chore that didn’t need 20 agents
The research brief for this post suggested a real fleet chore as the demo: audit every skill file for Title Case Airtable field names written to my Keywords table. That table is the only snake_case one in the base, and in October that bug made a whole set of status updates fail silently. It’s a textbook “fan out one agent per file” job, the same shape as the docs’ own audit example.
So I wrote it as an 18-line Python script instead. It scanned the 39 live skill files in 0.01 seconds and found 0 Title Case write payloads aimed at the Keywords table. To prove it wasn’t just blind, I pointed it at the 38 pre-fix backup files. It flagged 11 bad fields across 3 of them: the old blog-writer, keyword-researcher and content-researcher. Cost: zero tokens.
That’s the honest lesson. A workflow earns its keep when each item needs judgment: reading a route handler and deciding whether the auth check is real. When the rule fits in a regex, the script wins on speed, cost and repeatability, and it can run inside the cron job forever.
What Ultracode Costs, and the Guardrails

There’s no per-run price tag. Workflows simply spend tokens from many agents at once, and on a subscription every one of those tokens comes out of your session and weekly limits. The docs say a session with ultracode on “reaches a session or weekly limit sooner than the same work with it off.” Workflow agents run on your session model unless the script routes a stage elsewhere, so check /model before you start something big.
The guardrails that actually exist:
- Size guideline (advice, not a cap):
small<5 agents,medium<10 (the default),large<50, orunrestricted. Pro defaults tosmallon v2.1.271+. Set it with/config workflowSizeGuideline=smallor theworkflowSizeGuidelinesettings key. - Hard runtime caps: 16 concurrent agents by default (1–256 via
CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS, v2.1.269+), 4,096 items perparallel()/pipeline()call, 1,000 agents per run. - The
Large workflowwarning fires above 25 agents or 1.5M projected tokens. It’s advisory only, and it is switched off while ultracode is on, because the toggle counts as opting in to large runs. - Cache TTL: workflow agents get a 5-minute prompt cache by default, even on a subscription.
subagentPromptCacheTtl: "1h"extends it, and on the API that costs more per write. - No token cap. There’s an agent cap, not a spend cap. In headless mode,
--max-budget-usdis a client-side estimate, and my dispatcher doesn’t set it yet. That’s on my list, and I flagged it in the cheat sheet too.
The docs’ own advice is the right one: run the workflow on a small slice first (one directory, one narrow question), watch per-agent token use in /workflows, and stop it there if it balloons. Completed agents keep their results when you relaunch.
When to Use Ultracode, and When to Leave It Off

Here’s the rule I’d hand anyone, whether you’re at a laptop or running a fleet.
Use the keyword (one task) when…
- The job is bigger than one context window: a 500-file migration, a codebase-wide bug sweep, research that needs sources cross-checked against each other.
- Each item needs judgment, and you want independent agents to adversarially verify each other’s findings before anything reaches you. That quality pattern is the real reason workflows exist, more than raw speed.
- You’ll run it again. Press
sin/workflowsto save it as/your-command, then call it by name next time.
Use the toggle (whole session) when…
- You’re starting a big, unfamiliar piece of work and you’re watching. Then turn it off with
/effort ultracode offwhen you go back to routine edits.
Leave it off when…
- The check fits in a script. See the 0.01-second audit above.
- You’re on Pro and you want the week’s quota to last the week.
- The session is unattended. No human means no pause at the usage limit, no approval prompt in bypass mode, and failed agents can turn into a “successful” run with holes in it. In my fleet each skill is one focused job of about 4–11 minutes. If any of them ever needs a workflow, it’ll be a saved workflow approved by name with
Workflow(<name>), not a session-wide toggle.
If you’re trying to work out which of your own jobs belong in a script, which in a single agent, and which really justify a fan-out, that’s exactly the call I help operators make. Book an automation strategy session and bring your messiest workflow.
For the primitives underneath, see agent teams (a lead agent supervising peers, rather than a script) and what actually runs in production across my brands.
Claude Code Ultracode FAQ

Is ultracode an effort level?
No. It sits in the /effort menu, but the docs call it “a Claude Code setting rather than a model effort level.” It decides whether Claude orchestrates a workflow. Effort decides how hard each agent reasons.
Does ultracode work on the Pro plan?
Yes, once you switch Dynamic workflows on in /config. Pro defaults to the small size guideline (fewer than 5 agents) on v2.1.271+, and every agent draws on your usage limits.
Does it work with Sonnet?
On Sonnet 5.5, yes, because it supports xhigh. On Sonnet 4.6 and Opus 4.6, no: those models top out below xhigh, so ultracode isn’t offered.
Is ultracode the same as ultraplan or ultrareview?
No. /ultraplan has been retired (it’s in the removed-commands table of my cheat sheet), and planning mode is its own thing. claude ultrareview is a cloud-hosted multi-agent code review of your branch or PR. It’s related in spirit, but it’s a separate command. See my Claude Code review write-up.
Why didn’t typing “ultracode” in my script or cron job do anything?
Because the keyword only counts when a human types it. Prompts passed with -p, unstamped SDK prompts, scheduled-task prompts, webhooks and PR comments don’t trigger it. To allow a workflow headless, use a Workflow allow rule, auto mode, bypass mode or a hook, or put the ultracode setting in that session’s settings.
Can I stop a runaway workflow?
Yes. Open /workflows, select the run and press x. Completed agents keep their saved results if you relaunch in the same session. In a -p run there’s no /workflows screen to watch, which is one more reason to keep the size guideline small and use saved, named workflows there.
Final Thoughts
Claude Code ultracode is a genuinely useful switch. It’s also easy to misread. The keyword runs one task as a fan-out of agents. The toggle hands Claude the decision for a whole session. Neither one makes the model smarter. They make it wider.
For interactive work on hard, judgment-heavy jobs, try the keyword on a small slice and watch /workflows. For unattended agents, remember the asymmetry: the keyword can’t fire under -p, but the setting can, bypass mode won’t ask, and an agent that hits the usage limit fails instead of waiting. Decide on purpose, write the cheap script first, and save the fan-out for the jobs that really need a second opinion.

Get the AI Playbook
Every week: one real job I hand to an agent, what it costs, and what broke. No fluff, unsubscribe anytime.

📥 FREE: THE AI PLAYBOOK
The exact tools and workflows I use to run a one-person agency. 25 years of marketing experience distilled into an actionable guide. Yours free.
