A dynamic workflow is Claude writing a small program that runs a whole team of subagents for you, dozens or sometimes hundreds of them, in the background while your chat stays free to keep working. What makes it different from plain subagents or agent teams is where the plan lives. Normally Claude is the one deciding what to do next, turn by turn, and everything its helpers dig up gets dumped back into your single conversation. With a workflow, the plan moves into the program itself, a script in Claude’s terms, which holds the loop, the branching, and all the messy in-between results, so your chat only ever sees the finished answer, and that’s what lets it coordinate far more agents than one conversation could ever keep straight.

THREE WAYS TO START A DYNAMIC WORKFLOW

You never write the program yourself, since Claude builds it from whatever you describe in plain English, and there are three ways to kick one off, from gentlest to most all-in. The first, and the easiest to try, is the built-in /deep-research command, which is the one workflow that ships ready to go: you give it a question and it fans searches out across several angles, cross-checks the sources against each other, and hands back a cited report with the shaky claims already filtered out.

/deep-research What are the real tradeoffs between Vercel and Netlify for a small static blog in 2026?

If it feels like Claude must have already been doing this whenever you asked it to research something, it wasn’t: a normal research request is one Claude running a handful of searches inside your chat and handing you its confident best guess, so the job of catching hallucinated sources and stretched claims always fell on you. /deep-research flips that around by sending out a coordinated team that works in the background and only reports back once the claims have been checked against each other, so a lot of that validation work happens before the answer ever reaches you.

The second way is for a one-off job that’s bigger than a normal prompt, like an audit, a sweep, or a review across your whole project. Just ask for one, since saying “use a dynamic workflow” in the request is all it takes, and Claude builds the program around that single task instead of grinding through it itself. The keyword ultracode isn’t a separate feature, it’s just the one-word version of that same ask, handy once you know you want the token-heavy treatment and don’t feel like spelling it out.

ultracode: review everything that changed on this branch and flag anything that could break before I merge it

WHAT IS ULTRACODE A keyword Claude Code watches for: dropping it anywhere in a prompt tells Claude to run the task as a full background workflow, agents and all, without you having to describe any of the machinery. Not to be confused with ultrathink, which just makes Claude think harder on a single normal reply.

The third way is the all-in option. If you’re settling in for a session of heavy lifting, flip /effort ultracode and Claude runs a workflow for every substantial task without being asked, then switch back to /effort high when the heavy stretch is over.

WHY MORE AGENTS MEANS BETTER ANSWERS, NOT JUST FASTER ONES

The point isn’t that more agents are faster, it’s that they check each other. The script delivers what Anthropic calls “a repeatable quality pattern, not just run more agents”: independent agents adversarially reviewing each other’s findings, or drafting a plan from several angles and weighing them against each other, so instead of one confident pass you have to take on faith, you get a result where the weak claims already got argued down before you ever see them. The checking only counts if the agents are actually independent: a single reviewer tends to find one plausible answer and stop looking, and once it commits to a theory, everything it does afterward gets quietly bent to fit it, but agents poking holes in each other’s work never get the chance to anchor on the same first guess. In Anthropic’s testing, a lead Claude directing a team of subagents beat a single Claude working alone by 90.2% on their internal research eval.

There’s a second payoff that’s easy to miss: because the plan is written down as a script, you can save a run that did exactly what you wanted and rerun the same orchestration later. Saving happens right in the /workflows view, where pressing s turns the run into a slash command of your own, so a branch review you liked becomes /review-branch and gets the same scrutiny each time instead of being reinvented from scratch on every run.

HOW DYNAMIC WORKFLOWS COMPARE TO MULTI-CLAUDING, SUBAGENTS, AND AGENT TEAMS

These all sound similar, so here’s the quick way to tell them apart, from most hands-on to most automated:

  • Multi-clauding is you running several separate Claude sessions yourself, with no coordination between them, just parallel tracks you manage by hand.
  • Subagents are a few helpers that go off, do one thing, and report back without ever talking to each other, which makes them great for independent side tasks like research or codebase exploration.
  • Agent teams are a small crew, usually three to five, that actually coordinate through a shared task list and a team lead, for work where the agents need to know what the others are doing.
  • Dynamic workflows are a script Claude writes that runs the whole show, dozens to hundreds of agents, cross-checks them, and can be saved and run again. Use one when the job outgrows a handful of subagents, when you want the findings verified against each other, or when you want to save the whole process and run it again later.

PRACTICAL USES FOR A BLOG OR PRODUCT: AUDITS, SWEEPS, AND COMPARISONS

You don’t need a 500-file migration to make this useful. The everyday wins for someone running a site or a product are sweeps and comparisons that would be tedious to babysit one at a time, and the two I run most often on this blog are exactly that: a broken-link audit, where every post gets checked and I get back a list of dead internal links grouped by file, and a voice-and-tone sweep, where every article gets read against my style guide and the sentences that break my rules come back flagged. An SEO pass rounds out my regular rotation, checking every post’s frontmatter for weak excerpts, missing tags, titles that break my casing rules, and dates in the wrong format, then handing me a fix-list per file.

ultracode: read every article in src/content/posts and flag sentences that break the voice rules in my content guide, grouped by file

The comparison side is just as handy. You can have Claude draft three genuinely different homepage approaches and then have separate reviewers weigh them against each other and recommend one with reasons, which beats asking a single Claude to pick its own favorite.

TOKEN COSTS, USAGE LIMITS, AND OTHER GOTCHAS

Dynamic workflows can chew through your usage alarmingly fast. Anthropic has measured multi-agent systems at roughly fifteen times the tokens of a regular conversation, and it all counts against the same limits as your everyday coding. When the feature launched, the story making the rounds was someone who spun up 62 subagents and maxed out their five-hour usage cap in eighteen minutes, so a few habits are worth building early.

Pilot on a small slice first: run one folder or one narrow question, watch the per-agent token count in the /workflows view, and use that to estimate the full job, since you can stop a run at any point without losing finished work. Check /model too, because every agent runs on your session’s model, so a hundred-agent run on Opus pays Opus prices a hundred times, though you can tell Claude to use a cheaper model for the grunt stages.

One more thing that surprises people: workflow agents approve their own file edits no matter what permission mode you’re in, so for anything that touches your code, run a read-only “find the problems” pass first, read the results yourself, and only then run a pass that makes the changes. A bigger team makes the work go faster, not more correct, and a tidy green checkmark at the end still isn’t a reason to merge without reading.

WHEN A DYNAMIC WORKFLOW IS OVERKILL

Most tasks don’t need any of this. A quick lookup or a routine edit is cheaper and faster as a plain prompt or a single subagent, and workflows have two limits worth knowing: a run can’t stop to ask you anything, so work that needs your input between steps has to be split into separate runs, and a run only survives inside the session that started it, so quitting Claude Code means starting over. Try out /deep-research to get familiar with how a run feels, and go from there once you hit a job genuinely too big for one conversation to hold.