Workflows, Deep Research & Ultracode: When and How to Use Them

A practical guide to Claude Code's new orchestration features (Opus 4.8)

Quick clarification before the detail: "workflows" and "dynamic workflows" are the same feature — "dynamic workflow" is the full name, "workflow" is the shorthand. Deep research isn't separate either; it's one pre-built workflow that ships in the box. So there are really three things to understand: the workflow system, the deep-research workflow as a ready-made example, and ultracode as the switch that fires workflows automatically.

1. Workflows (a.k.a. Dynamic Workflows — same thing)

What it actually is, mechanically. Normally Claude holds the plan in its own head (its context window) and decides turn-by-turn what to do next. A workflow flips that: Claude writes a small JavaScript program that holds the plan — the loops, the branching, the list of files, the intermediate results — and a separate runtime executes that program in the background, spawning subagents to do the actual work. The key consequence: your context window only ever sees the final answer, not the hundred messy steps. That's why it scales to hundreds of agents without Claude "forgetting" halfway through.

When specifically to use it

• The job needs more agents than one conversation can juggle — a codebase-wide bug sweep, a 500-file migration.
• You want the orchestration saved and rerun — e.g. the exact same review on every branch.
• You want a quality pattern, not just more workers — independent agents adversarially checking each other's findings, or drafting a plan from several angles and weighing them. A single Claude pass can't cross-examine itself; a workflow can.

Best use cases: large migrations, security/auth audits across many endpoints, "find every place we do X" sweeps, cross-checked research, generating N independent solution drafts and picking the best.

How to use it — three ways

1. Put the word "workflow" in your prompt. e.g. Run a workflow to audit every API endpoint under src/routes/ for missing auth checks. Claude highlights the word, writes a script, and asks you to approve before running. Press alt+w to cancel the trigger if you didn't mean it.
2. Let ultracode decide (see section 4) — automatic.
3. Run a saved one as /<name>.

You approve once (the prompt shows the planned phases; pick "Yes, run it"). Watch it with /workflows. Save a good run as a reusable command by pressing s in that view (saves to .claude/workflows/ for the repo or ~/.claude/workflows/ for every project).

Pros

• Handles scale a single chat can't (up to 1,000 agents/run, 16 at once).
• Session stays responsive — it runs in the background.
• Repeatable and readable — it's a script you can inspect and rerun.
• Resumable mid-run (within the same session).
• Built-in quality patterns give more trustworthy output than one pass.

Cons

• Token cost — many agents means meaningfully more tokens than a normal chat for the same task. Counts against your plan limits.
• No mid-run input — you can't answer a question halfway. For sign-off between stages you must split into separate workflows.
• Resume dies if you quit Claude Code — a workflow in flight starts fresh next session.
• Overkill for small jobs — if one chat can do it, a workflow just burns tokens.
• Research preview, so expect rough edges.

2. Deep Research

What it is. A bundled, pre-written workflow — you don't author it, it ships ready. It fans web searches across several angles on your question, fetches the sources, has agents cross-check claims against each other and vote, then returns one cited report with the claims that failed cross-checking already stripped out.

When specifically to use it: a research question where you care about accuracy, not just a quick answer — anything where a single web search could hand you something plausible but wrong. "What changed in X between v20 and v22", competitive teardowns, "is this claim actually true." Use it when you'd otherwise open ten tabs and try to reconcile them yourself.

Best use cases: technical fact-finding, market/competitor research, verifying a contested claim, anything you'd want footnotes for.

How to use it — one command

/deep-research <your question>

Requires the WebSearch tool to be available. Approve it, then watch via /workflows or read the report when it lands.

Pros

• Cross-checking + voting filters out hallucinated/unsupported claims — far more trustworthy than a one-shot search.
• Cited, so you can verify each claim.
• Runs in the background; you get a clean report, not a transcript.

Cons

• Slower and more token-heavy than just asking.
• Web-only — it researches sources, it won't reason over your private files unless you point agents at them.
• Same preview caveats as workflows.

3. The contrast: Subagents vs Skills vs Workflows

The point of workflows only lands if you know what they replace. Same job, different "who holds the plan":

Rule of thumb: a few delegated tasks → subagents. A repeatable procedure you follow → a skill. A job too big for one context, or one you want codified and rerun, or one needing cross-checking → a workflow.

4. Ultracode

What it is. An effort setting that combines xhigh reasoning (Claude thinks as hard as it can) with automatic workflow orchestration (Claude decides on its own when a task deserves a workflow, instead of waiting for you to type "workflow"). One request can become several workflows in a row — one to understand the code, one to make the change, one to verify it.

When specifically to use it: high-stakes, substantial work where you want maximum thoroughness and don't care about cost or speed — a gnarly migration, a critical bug you can't afford to get wrong, a deep audit. Not for routine work.

How to use it — one command, then it's automatic

/effort ultracode

After that, Claude auto-plans workflows for every substantial task for the rest of the session. It resets when you start a new session. Drop back to normal with /effort high when you return to routine work. Only available on models that support xhigh effort (Opus 4.8 does).

Pros

• Hands-off — you stop having to remember to type "workflow".
• Maximum reasoning depth + auto-verification on everything substantive.
• In Auto permission mode it even skips the per-run approval prompt.

Cons

• Every task costs more and takes longer — it applies session-wide, including tasks that didn't need it.
• Easy to leave on by accident and burn tokens on trivial work (check /model and your effort before a big session).
• It decides when to fan out, so you cede some control over spend.

The one-line decision guide

• Quick answer or small task → just chat. No workflow.
• A few sub-tasks this turn → subagents (automatic, Claude does it).
• Researching a question you need to trust → /deep-research <question>.
• Big / repeatable / needs-cross-checking job → type "workflow" in your prompt.
• A whole session of high-stakes work, cost no object → /effort ultracode, then /effort high when done.

Limits worth remembering

• Up to 16 concurrent agents, 1,000 total per run.
• No mid-run user input; only agent permission prompts can pause a run.
• The workflow script itself has no filesystem/shell access — the agents read, write, and run commands; the script only coordinates them.
• Subagents always run in acceptEdits mode and inherit your tool allowlist. Add commands the agents need to your allowlist before a long run to avoid mid-run prompts.
• Turn workflows off via /config (Dynamic workflows toggle), "disableWorkflows": true in settings.json, or CLAUDE_CODE_DISABLE_WORKFLOWS=1.

Requires Claude Code v2.1.154+. Available on all paid plans; on Pro, enable from the Dynamic workflows row in /config.