Claude Code — workflows
Advanced Dynamic workflows, fan-out, verification, and how ultracode orchestrates them (2026).Parent map: Claude Code 2026.
Official: Orchestrate subagents at scale with dynamic workflows
What is a dynamic workflow?
Instead of you manually spawning every worker, Claude can:
- Plan a multi-step approach for a substantive task
- Fan out work to subagents (sometimes many)
- Collect results
- Optionally grade / revise (performance outcomes — worker revises until a rubric is met)
Anthropic’s 2026 direction: lead agent plans dynamically and can scale parallel subagents in one session, with stronger verification loops.
Ultracode ↔ workflows
Ultracode is a Claude Code effort setting, not a separate binary:
- Sends xhigh reasoning effort to the model
- Turns on automatic workflow orchestration for substantive tasks
# Start a session already in ultracode
claude --effort ultracode
# Inside a session
/effort ultracodeOne-shot workflow without changing session effort (keyword in the prompt):
ultracode: audit every API endpoint under src/routes/ for missing auth checksExplicit orchestration:
create a workflow that migrates every internal fetch() call to the new HttpClient wrapper
use a workflow to migrate every component under src/components/ from styled-components to Tailwind,
working on each file in its own isolated copyWith ultracode on, one user request may become several workflows in a row (understand → change → verify). That uses more tokens and time than /effort high.
Drop back:
/effort highRequires a recent CLI (docs cite v2.1.203+ for ultracode) and a model that supports xhigh effort. Confirm with your claude --version and model config.
Details: claude-code-ultra.
Common workflow shapes
1. Explore → Plan → Execute → Verify
Explore: read-only subagents map modules
Plan: /plan or plan permission mode; you approve
Execute: implement on a branch / worktree
Verify: tests + /code-review (+ optional second agent)2. Fan-out by area
Lead: "Dynamic workflow: implement feature X.
Parallelize independent files; one worker per area;
return diffs only; main session integrates."3. Batch across the codebase
/batchDecomposes a large change into independent units; often pairs with worktrees.
4. Graded workers (when available)
Lead defines success criteria; each subagent result is scored; failures revise until they pass — fewer silent half-done workers. See workflows docs for current “performance outcomes” behavior.
Settings & scale (high level)
Production setups in mid-2026 often combine:
| Knob | Role |
|---|---|
| Fallback model chains | Survive overloads |
| Nested subagents | Depth of delegation (can get expensive) |
| Max parallel workers | Cap fan-out |
| Hooks | SubagentStop, permission gates |
| CLAUDE.md rules | When to workflow vs stay single-threaded |
Prefer project settings + documented rules over ad-hoc “spin 50 agents.”
Example shape (illustrative — keys change; verify against settings):
{
"permissions": {
"defaultMode": "acceptEdits"
}
}Use official docs for nested-agent depth, concurrency caps, and fallback chains — wrong configs can burn quotas quickly.
Everyday vs workflow mode
| Routine | Substantive |
|---|---|
| Typo, one-file fix | Feature spanning many files |
| Quick question | Unfamiliar subsystem |
/effort high or default | /effort ultracode or explicit workflow prompt |
| Single session | Explore + plan + parallel execute |
Cost discipline
- Cap concurrent workers
- Prefer short summaries from workers
- Use
/costand account dashboards - Ultracode = session-scoped; reset for small tasks
- Don’t combine ultracode + max parallelism + cloud ultra review unless needed