Claude blog · Explainer

Dynamic workflows in Claude Code, explained simply

Claude can now build its own team of helpers for a task instead of doing everything by itself in one go. Here is the plain-English version of the Claude blog post.

June 2, 2026  ·  Source: claude.com/blog, “A harness for every task”

Normally Claude works on your task alone, in one long conversation. A workflow lets Claude split the job across many separate Claudes, each handed one small piece, then combine their answers. Like a manager assigning work to a team instead of doing all of it personally.

When one Claude does a big job in a single sitting, three things go wrong. Splitting the work into fresh, separate helpers fixes all three.

It quits early Does part of the job, then declares it finished. A fresh helper assigned just that piece actually completes it.
It grades its own homework When asked to double-check its own work, it tends to agree with itself. A separate helper checks with no ego in the game.
It drifts off-target Over a long back-and-forth it slowly loses the plot. A helper with one short, clear instruction stays on point.

Claude writes a short script that hands out the work, collects it, and stitches it together. If it gets interrupted, it picks up where it left off instead of starting over.

One big task Split into pieces Many helpers work at once Check each other Combine into one answer

A handful of repeatable shapes do most of the work.

Sort and route. Look at what came in, send each item to the right specialist.
Split and merge. Break a job into parallel parts, then combine the results.
Check each other. One helper makes a claim, another tries to prove it wrong before it counts.
Make many, keep the best. Generate lots of options, filter down to the good ones.
Run a tournament. Compare options head-to-head until a winner stands.
Keep going until done. Spawn more helpers until nothing new turns up.

The blog is blunt about this: workflows are not for every task. A team of helpers burns a lot more tokens than one Claude, so it is worth it only for big, high-value jobs.

Good fit. Complex, high-stakes work where being thorough and double-checked is worth the extra cost: large reviews, deep research, ranking or sorting a big pile, anything you would normally split across a team.
Bad fit. Routine, quick tasks. The blog warns it “may end up using significantly more tokens,” so reaching for it on small jobs is wasteful.
How to trigger it. Just ask for it, or say the word ultracode in your message to push Claude Code to build one.
Keep a lid on cost. You can set a budget in plain language, like “use 10k tokens,” and it works within that.

The part the author is most excited about: this is not just for coding. Real non-technical jobs it can run:

Rank 80 ad creatives for one campaign, then double-check the top ten before you trust them.
Pressure-test a plan from three angles at once: investor, customer, competitor.
Read a quarter of workflow reports and pull out the patterns that keep repeating.
Brainstorm names for a campaign or tool, then run a tournament to pick the top three.
Mine past sessions for the corrections you keep making, and turn them into standing rules.
Auto-sort and de-duplicate a pile of support or feedback tickets.

Claude blog, “A harness for every task: dynamic workflows in Claude Code” · claude.com/blog
Shared by Thariq (@trq212), June 2, 2026 · x.com/trq212