Workflows

Dynamic workflows (research preview, shipped alongside Opus 4.8 on 2026-05-28) let you hand Claude a large, decomposable job and have it write an orchestration script that fans out across tens to hundreds of subagents running in the background, checking its own work before results reach you. It's the tier above ad-hoc delegation: where a single subagent is a one-shot handoff and Agent Teams are peer sessions that message each other, a workflow is deterministic, scripted fan-out with built-in verification and saved progress.

There are two ways in. For one task, put the keyword ultracode in your prompt (or just ask for a workflow in plain words); the session's effort level stays as it is. For a whole session, turn on the Ultracode toggle: /effort ultracode, or Tab in the /effort slider. Since 2.1.284 the toggle is separate from the effort level: it no longer forces xhigh, and it stays on at any level until you run /effort ultracode off. With it on, Claude decides which substantive tasks warrant a workflow. Either way, you watch live runs with /workflows.

How you use it day-to-day

• Reach for it on work one context can't hold. Codebase-wide migrations, broad audits, multi-file refactors, multi-source research: anything that decomposes into many similar independent units.
• Works beyond code. The same fan-out → verify → synthesize structure works on non-engineering jobs too: multi-section reports, research digests, content or ops batches that split into many similar units. The Claude Code team frames workflows as the biggest capability jump since skills and subagents precisely because the units don't have to be code: describe a decomposable knowledge task, say ultracode, and the orchestration is identical.
• Let Claude write the orchestration. Describe the goal and say ultracode; it builds the fan-out → verify → synthesize structure and runs it in the background while you keep working.
• Watch it with /workflows. The run executes detached and re-engages you when it finishes; /workflows is the live progress view.
• Isolate file-mutating agents. Workflow agents can run in their own git worktree so parallel edits don't collide. Use that when the agents write files.

Gotchas

• Research preview. The surface, keywords, and behavior are still moving; check the changelog before scripting against it.

• Cost scales with fan-out. Tens to hundreds of agents is real token spend. Use workflows when the scale justifies it, not for routine edits. A single subagent or a quick edit is usually the right tool.

• Distinct from Subagents and Agent Teams. Subagents = one-shot delegates; Agent Teams = peer sessions; Workflows = scripted, verified orchestration on top. Don't conflate them.

• The session toggle spends more. With Ultracode on, one request can become several workflows in a row (understand, change, verify), so each task uses more tokens and takes longer. On a subscription it draws down your limits faster. Turn it on for a stretch of hard work, then off.

Keyword for one task, toggle for a session. Without either, Claude won't author a workflow; it does the work inline or delegates a subagent or two. Typed ultracode by mistake? Option+W (Alt+W) dismisses the highlight for that prompt.