Analysis
If you've ever watched a job sit in a queue while one process grinds through it one step at a time, you already understand the problem Anthropic is going after.
On 28 May 2026, alongside Claude Opus 4.8, Anthropic shipped a Claude Code feature called Dynamic Workflows (Anthropic, Introducing dynamic workflows in Claude Code). The pitch is straightforward. Hand Claude a large, messy task, and instead of working through it alone, it writes a short program that splits the work up, hands each piece to a separate Claude "subagent," and runs them at the same time. Then it checks the pieces and combines them into one answer.
The headline demo Anthropic put forward wasn't a slide of benchmarks. It was a real port of the Bun JavaScript runtime's code from Zig to Rust: roughly 750,000 lines, a 99.8% pass rate against the existing test suite, and eleven days from first commit to merge (MarkTechPost, Bun's Zig-to-Rust port). That's the kind of job that would normally tie up a senior team for weeks.
For Australian business teams, the "so what" is this: work that used to be done one slow step at a time can now be fanned out across many workers at once. That can turn an afternoon into a coffee break. It can also burn through a lot more tokens than a normal session, so it's worth understanding before you point it at your codebase.
How It Works
Dynamic Workflows runs in three rough stages: breaking the task down, doing the work, and pulling it back together.
First, the breakdown. Claude reads the task and splits it into smaller subtasks. You can do this yourself by spelling out what each subagent should handle, or you can let Claude plan it. The automatic mode earns its keep on open-ended work, where you don't know the best way to carve up the job until you've started (MarkTechPost, Dynamic Workflows activation and behaviour). It kicks in when you put "workflow" in a prompt, switch on the "ultracode" setting, or run the bundled /deep-research workflow.
Worth being clear on the mechanics, because the framing matters: this isn't a generic developer API you call from your own app. It's a Claude Code feature. Claude writes a JavaScript script that orchestrates the subagents, and that script runs inside Claude Code on the CLI, Desktop, or VS Code. The script itself can't touch the filesystem or shell; only the agents can.
Second, the work. Each subtask goes to a Claude instance with the right tools for the job, and the subagents run side by side rather than one after another. Anthropic caps this: up to 16 agents running concurrently, and up to 1,000 agents in total across a single run (MarkTechPost, concurrency limits). Reports also describe subagents being able to assign work to a model that suits the subtask, though the exact per-subagent model logic isn't documented. One detail floated in early write-ups, but not confirmed by Anthropic, is a shared store where subagents drop and pick up intermediate results; treat that as unconfirmed for now.
Third, the synthesis. The outputs from all the subagents get combined into the final result. Anthropic's account also describes a refutation-and-iteration step and a verification pass before anything is returned, so the system isn't just gluing fragments together; it's checking them. How the merge happens depends on the task and the output you need.
Use Cases and Performance
Dynamic Workflows shines on work that splits cleanly. Reviewing a large codebase divides by file or module, with a subagent on each piece. Research synthesis splits by source or topic. Data analysis splits by partition. Support triage splits by ticket type. If a task naturally breaks into chunks that don't depend on each other, this is where it fits.
On speed, the honest version is hedged. Anthropic hasn't published a tidy benchmark table for these scenarios. Early secondary coverage offered an illustrative estimate that a research-and-synthesis job might drop from roughly 40 minutes to somewhere around 8 to 12 minutes (MarkTechPost, research/synthesis estimate). Specific figures that circulated for other tasks (for example, a 45-minute security audit finishing in under 3 minutes, or a 2-hour synthesis finishing in 8) aren't backed by any Anthropic source, so don't bank on them. As a rule of thumb, more agents means more speed on parallel work, up to the point where coordinating them starts eating the gains.
Cost is the part to watch. Anthropic's own warning is blunt: a dynamic workflow can chew through far more tokens than a typical Claude Code session. You're paying for every subagent's work. There's a tempting line of reasoning that running 50 agents in parallel costs the same as 50 sequential requests, no discount and no penalty, but Anthropic hasn't confirmed that pricing, so it's better read as plausible than settled. The trade is speed for spend. If you need the answer fast and the budget can wear it, parallel is the move. If cost is the binding constraint, doing it sequentially is cheaper.
The New Compute Paradigm
There's a real shift in how AI work gets done here. For most of the short history of these tools, inference has been sequential: one prompt, one response. Even a long back-and-forth is still a chain, each turn waiting on the last. Dynamic Workflows runs many inference jobs at once and combines what comes back.
That changes a few things downstream. Providers have to support a lot of concurrent inference without latency falling apart. Pricing starts to think in terms of a whole "workflow" rather than a single request, with the bill driven by how many subagents you run, how long they take, and what the merge costs. And capability-wise, a team of agents can chew through problems that won't fit in one model's context window or that strain a single model's reasoning.