Claude Code now spins up dozens to hundreds of subagents in parallel that verify each other adversarially. The acid test: porting roughly 750k lines of the Bun runtime (from Zig to Rust) in 11 days.
Think about the last time you faced a huge, tedious task: renaming hundreds of files, combing through an entire project hunting for bugs, migrating old code. You did it yourself, by hand, cursing under your breath, right? We've all been there. Well, the leap for coding agents is no longer "write me this function," it's "orchestrate an army to do the heavy lifting for me." Anthropic has made Dynamic Workflows in Claude Code generally available, and the number they dropped commands respect: anywhere from dozens to hundreds of subagents working in parallel within a single session. It's no longer an assistant, it's an entire brigade at your command.
An agent that trips itself up
Here's the genuinely clever part, and it's more than it looks. Once triggered, Claude plans based on your request, breaks the work into subtasks, and distributes them among agents that run simultaneously. But the magic trick is adversarial verification: some agents attack the problem from independent angles while others actively try to refute what the first ones found, and the whole thing iterates until they all converge on the same answer. It's like having a team where someone always plays devil's advocate so no error slips through. And as if that weren't enough, progress saves itself: if something gets interrupted midway, it resumes where it left off instead of starting from scratch — which is exactly the nightmare we all dread.
The example that makes it all click
Lab figures are all well and good, but nothing convinces like a killer example. The flagship demonstration was the rewrite of Bun, an entire runtime: roughly 750,000 lines of Rust ported from Zig, completed in eleven days from the first commit to the final merge. Stop and think about that for a second — a project that would have taken a human team months, solved in under two weeks by an orchestra of agents. Anthropic lays the obvious use cases on the table: code audits, bug hunts, security reviews, and migrations across thousands of files. Exactly the kind of work we hate doing by hand and that, conveniently, AI does without complaining.
For NeuralOS this is doubly relevant
And here we're not going to play modest, because this hits us head-on on two fronts. On one hand, we use exactly this kind of multi-agent orchestration in our own build-and-audit workflow — in fact, much of what you see in NeuralOS was built and reviewed this way, with armies of agents verifying each other. On the other hand, it's the definitive proof that "one builder plus many agents in parallel" has stopped being a pretty demo and become a real way to work in production. That's precisely the power we want to put in your hands: not that you build faster on your own, but that you direct a team of agents that works while you think about the next idea.