Claude Opus 4.8 with UltraCode lets one instruction run hundreds of self-checking agents in parallel.
Opus 4.8 shipped with a new ability called dynamic workflows, switched on by a setting called UltraCode. Instead of one helper doing one thing at a time, Claude breaks a big job into pieces, hands each piece to its own helper, runs hundreds of them at the same time, then has other helpers double check the work before you see it.
A developer used it to rebuild a huge software project from one programming language into another, around 750,000 lines of code. Hundreds of agents worked in parallel with two reviewers on every file, and it was done in 11 days, work that would normally take a human team months.
His honest take is to turn UltraCode on for the big jobs: reviews, giant cleanups, a plan you want stress tested. Leave it off for little quick chats. A swarm suits one goal at the top that splits down into sub-agents and subtasks; for something small, just use normal Claude Code.
If you are using the terminal directly there are two ways: turn on the Claude Code specific setting UltraCode to get started, or just ask Claude to create a workflow that is dynamic and it goes from there. In his own dashboard he prefers a toggle you flip on to wake a swarm.
Section three of the stack is verify: other helpers act like judges, trying to poke holes in the work and prove it wrong, so only what survives gets through. You get checked work, not guesses, and it saves a lot of time on quality control as well.
First is command: one plain instruction like audit my project, migrate this, or find every bug, with you as the conductor not playing every instrument. Then a swarm splits the job across hundreds of helpers, verify has judges poke holes, you watch it on a live map, and keep stores everything in one workspace.
He answers that you use a swarm only for big tasks, where there is one goal at the top that needs splitting into sub-agents and subtasks. A small task will not spawn that many helpers anyway, because that would use a lot of tokens; for small things just use normal Claude Code.
No. Some people say this sounds like it is only for coders, but it is not: you type what you want in plain words, give a short mission, and the swarm does the heavy lifting. You just read the result at the end.
One thing he always hears is that AI sounds confident even when it is wrong, which is true. Opus 4.8 is already far more honest about flagging its own mistakes, and the swarm doubles down on that, so nothing reaches you until other agents have tried to break it first.
It starts planning the workflow first, and that might take a couple of minutes to run before anything else. You can then see the full plan, including research it has done out on the web with the sub-agent team, and once that completes you watch the agents being set up underneath it.
“Claude Opus 4.8 just dropped and Anthropic also just dropped UltraCode and it's the biggest AI agent upgrade I've seen all year.”
“Basically, you give a command and hundreds of agents work together as a swarm.”
“you're basically the conductor so you don't play every instrument”
“so only what survives gets through. And you get checked work, not guesses.”
“now, whilst this is brand new, are going to be miles ahead when everyone else catches up.”
“Don't be most people, right?”
“Claude Opus 4.8 Ultracode is INSANE!.” https://www.youtube.com/watch?v=_lV7CB4NVCU. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/claude-opus-4-8-ultracode-is-insane.html.