Anthropic Brings Dynamic Workflows to Claude Code, Letting the Model Build Its Own Harness

Anthropic has introduced a feature it calls dynamic workflows in Claude Code, and the idea behind it marks a quiet but meaningful shift in how AI coding assistants approach hard problems. Instead of forcing every request through one predetermined sequence of steps, Claude can now author a workflow on the fly, essentially writing a small orchestration script that decides what runs in parallel, what runs in sequence, and how the results get checked. The company frames it with the phrase a harness for every task, capturing the core notion that no single fixed pipeline fits every job, so the model should build the scaffolding it needs in the moment.

The mechanics are more interesting than a simple speedup. A dynamic workflow lets Claude spawn dozens of subagents, each working on an independent slice of a larger problem, and then pipeline their output through stages without waiting for every agent to finish in lockstep. A code review, for example, might fan out across separate dimensions like correctness, security, and performance, with each finding then handed to a fresh set of skeptical verifiers whose job is to try to refute it before it survives into the final report. This adversarial structure is deliberate. Large models are prone to producing plausible but wrong conclusions, and routing claims through independent checks is one of the more reliable ways to catch them.

What makes the approach distinct from earlier agent frameworks is that the control flow is deterministic while the reasoning inside each step stays model-driven. The loops, conditionals, and fan-out are encoded in a script, so the orchestration behaves predictably and can even be resumed from where it left off, but the actual judgment at each node still comes from Claude. That separation matters for the kinds of work that simply do not fit in a single context window, such as large migrations, sweeping audits, or research that needs to pull from many sources at once. The model can decompose the task, cover it in parallel, and synthesize the pieces back into a coherent result.

The broader signal here is about where agentic coding tools are heading. Rather than positioning the AI as a single assistant answering one prompt at a time, Anthropic is treating it as something closer to an engineering manager that can decide when to delegate, how many workers to deploy, and what verification is worth the cost. Because these workflows can consume substantial compute, the company has built in explicit opt-in and budget controls so the scale stays under user direction. It is a pragmatic acknowledgment that more agents are not always better, but when a task genuinely demands breadth or confidence, giving the model the freedom to design its own harness may prove far more effective than any one-size-fits-all pipeline.