Anthropic Launches Dynamic Workflows in Claude Code, Hundreds of Subagents Now Run in Parallel

Anthropic this week rolled out a feature it calls dynamic workflows in Claude Code, and the framing is more ambitious than the modest name suggests. Until now, getting a coding assistant to tackle a large, sprawling task meant babysitting it through one step at a time, or hand-wiring a rigid pipeline that broke the moment the work didn't fit the mold. The new approach flips that around. Instead of a human scripting the orchestration, Claude writes the orchestration itself — a small program that decides what to fan out, what to run in parallel, and what to verify before committing to an answer.

The headline capability is scale. A single dynamic workflow can launch anywhere from a handful to several hundred subagents, each working on its own slice of the problem, all coordinated from one session. That matters because the hard part of large software work is rarely any individual edit; it's holding the whole thing in your head at once. A migration that touches a thousand call sites, an audit that has to sweep every file, a review that needs independent perspectives on the same change — these are jobs that exceed what a single context window can carry. By decomposing the task and running the pieces concurrently, the workflow turns wall-clock time from the sum of every step into roughly the length of the slowest single chain.

What makes the design interesting is that it is deterministic where it needs to be and model-driven where that pays off. The control flow — the loops, the conditionals, the fan-out — lives in a script that runs predictably every time, while the judgment lives in the agents that script spawns. Anthropic has leaned into patterns that reflect how careful engineers actually work: adversarial verification, where several skeptical agents try to refute a finding before it survives; judge panels that generate competing approaches and score them against one another; and loops that keep hunting until a problem space comes up empty rather than stopping at an arbitrary count. The effect is less a faster autocomplete and more a small, self-organizing team.

The stated goal is blunt. Anthropic wants to compress work measured in quarters down to days, and dynamic workflows are the mechanism it is betting on. There are real questions still hanging over the approach — the token cost of running fleets of agents is substantial, and the company has been careful to gate the heaviest orchestration behind explicit user opt-in rather than letting it fire on every task. But the direction of travel is clear. The frontier of coding assistants is no longer about how good a single model is at writing one function; it's about how well that model can marshal many copies of itself to chew through problems no single pass could finish.