Claude Code Dynamic Workflows
Ultracode is the switch for Claude Code's dynamic workflows: write ultracode in your prompt (or just ask for a workflow in your own words) and Claude writes a script that orchestrates dozens to hundreds of subagents in the background. As of 2026-10-20 the trigger keyword has been renamed from workflow to ultracode, and the word workflow no longer triggers a run.
What is a dynamic workflow?
A dynamic workflow is a JavaScript script Claude writes for your task. A runtime executes it in the background while your session stays responsive, and only the final answer returns to the conversation.
A script, not a chat loop
Claude writes a JavaScript script that holds the loop, the branching, and the intermediate results. The orchestration itself is what gets repeated — not just the prompt.
Context stays light
Intermediate results live in script variables instead of Claude's context window, so the conversation only holds the final report — not every agent's output.
Built-in quality patterns
A workflow can have independent agents adversarially review each other's findings before reporting, or draft a plan from several angles and weigh them — more trustworthy than a single pass.
Read it, save it, rerun it
Because the plan is code, you can read the script before it runs, resume it after a pause, and save a good run as a /command you reuse on every branch.
Subagents vs Skills vs Workflows
All three run multi-step work. The difference is who holds the plan.
| Aspect | Subagents | Skills | Workflows |
|---|---|---|---|
| What it is | A worker Claude spawns | Instructions Claude follows | A script the runtime executes |
| Who decides what runs next | Claude, turn by turn | Claude, following the prompt | The script |
| Where intermediate results live | Claude's context window | Claude's context window | Script variables |
| Scale | A few delegated tasks per turn | Same as subagents | Dozens to hundreds of agents per run |
Three ways to start one
Available on all paid plans (turn it on in /config on Pro), plus the Claude API, Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry.
1. The ultracode keyword
Write ultracode in your prompt and Claude writes a workflow for the task instead of working turn by turn. Since 2.1.160 the trigger keyword was renamed from workflow to ultracode: the word workflow no longer triggers a run, while asking for one in your own words still works.
ultracode: audit every API endpoint under src/routes/ for missing auth checks2. /effort ultracode
/effort ultracode is now its own toggle (press Tab, or run /effort ultracode [on|off]): since 2.1.284 it no longer forces xhigh effort and stays on at any effort level, so Claude plans a workflow for every substantial task in the session.
/effort ultracode3. /deep-research
The bundled workflow: it fans out web searches, cross-checks sources, votes on each claim, and returns a cited report.
/deep-research What changed in
the Node permission model v20-v22?About /deep-research
/deep-research is the built-in workflow that investigates a question across many sources. It fetches and cross-checks what it finds, filters out claims that don't survive cross-checking, and returns one cited report. It needs the WebSearch tool enabled.
When to reach for a workflow
Use one when a task needs more agents than a single conversation can coordinate, or when you want the orchestration codified as a script you can rerun.
Codebase-wide bug sweeps — many finders in parallel, each verified before reporting.
Large migrations — a 500-file framework switch or language port, one agent per site.
Cross-checked research — sources weighed against each other, not a single summary.
Hard plans — drafted from several independent angles, then judged and synthesized.
Case study: rewriting Bun from Zig to Rust
Jarred Sumner used dynamic workflows to port the Bun runtime from Zig to Rust — a scale of change that would normally take a team weeks.
750K
lines of code
11 days
elapsed
99.8%
existing tests kept green
Limits worth knowing
The runtime applies a few guardrails (the values below follow the official docs as of 2026-10-20).
Up to 16 concurrent agents by default
Excess agents queue and start as slots open up; fewer run at once on machines with limited CPU cores. The default is up to 16 concurrent agents, and you can change it with CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS set to a value from 1 to 256, which requires Claude Code v2.1.269 or later.
1,000 agents per run
A backstop against runaway loops, set far above any normal workflow. Claude also gets an advisory size guideline when it writes a workflow: pick small (fewer than 5 agents), medium (fewer than 10), large (fewer than 50) or unrestricted in the Dynamic workflow size row in /config; the default is medium (small on a Claude Pro subscription since v2.1.271), the setting needs v2.1.219 or later, and it is advice rather than a hard cap.
No mid-run input
Only agent permission prompts and a usage-limit wait can pause a run. For sign-off between stages, run each stage as its own workflow.
Resumable in-session
Completed agents return cached results on resume; the rest run live. Exiting Claude Code restarts the workflow fresh. Since v2.1.271 a run can pause instead of failing the affected agents when they hit your claude.ai usage limit, wait for the reset and then continue on its own — but only in an interactive session signed in with a claude.ai subscription with autoContinueAtUsageLimit on, when the limit resets within 24 hours (a weekly limit can reset further out), and at most twice per run; otherwise the affected agent fails. claude -p, the Agent SDK and background sessions do not pause.
Workflows use more tokens — plan for it
A single run can spawn hundreds of agents, so it uses meaningfully more tokens than the same task in conversation, and counts toward your plan's usage and rate limits. The /workflows view shows each agent's token usage as the run progresses, and you can stop the run there at any time; the official guidance is to try a small slice first, one directory instead of the whole repo, before committing to a large task. Each agent uses your session's model unless the script routes a stage to a cheaper one. A session with ultracode on reaches a session or weekly limit sooner, because those tokens draw on your usage limits too.
Running big workflows through QCode
Because 1,000-agent runs burn tokens fast, the access tier you run them on matters. QCode provides higher-tier Claude access optimized for low latency in China, so codebase-scale workflows stay practical to run — no foreign card or phone number required. Not ready for a monthly plan? Pay-as-you-go Packs start at $4 for a one-time quota you use within 7 days. You can't buy packs while you have an active personal plan; if the plan's quota isn't enough, buy a booster pack.
Start building with Claude Code
Get Claude Code access through QCode and run dynamic workflows, ultracode, and Fable 5 / Opus 4.8 with China-optimized latency.