Claude Code's Dynamic Workflows let hundreds of AI agents tackle your biggest coding jobs in parallel
Anthropic's Dynamic Workflows for Claude Code orchestrate up to 1,000 subagents per run, handling massive migrations and audits that previously took weeks.
Anthropic has launched Dynamic Workflows in Claude Code as a research preview, and it represents a genuinely different approach to how AI handles large engineering work. This moves well beyond asking Claude to fix a bug or refactor a function. We are talking about running tens to hundreds of parallel subagents on a single task, with results that converge over hours, or even days, while your session stays completely free.
What Dynamic Workflows actually are
The simplest way to understand this: a dynamic workflow is a JavaScript script that Claude writes based on what you describe. A runtime then executes that script in the background, spinning up subagents to do the actual work. Claude’s context window holds the final answer, not every intermediate step along the way.
That last part matters more than it sounds. With traditional subagents, Claude acts as the orchestrator turn by turn, and every result lands back in its context, eating tokens and limiting how much total work can happen. With a dynamic workflow, the script holds the loop, the branching logic, and the intermediate state. Claude gets out of its own way.
Each run can use up to 16 concurrent agents, with a hard cap of 1,000 agents total. The workflow script itself cannot touch the filesystem or shell directly. Only the agents it spawns do that work, and they always run in acceptEdits mode, inheriting your existing tool allowlist.
The Bun port is the proof point
Jarred Sumner used dynamic workflows to port Bun from Zig to Rust. The end result: roughly 750,000 lines of Rust, 99.8% of the existing test suite passing, eleven days from first commit to merge.
The approach was straightforward in structure, if not in scale. One workflow mapped the correct Rust lifetime for every struct field in the Zig codebase. A second wrote every .rs file as a behaviour-identical port of its .zig counterpart, with hundreds of agents working in parallel and two reviewer agents on each file. A fix loop then drove the build and test suite until both ran clean. After the port landed, an overnight workflow identified unnecessary data copies and opened a PR for each one.
That kind of task would previously have taken a team of engineers several months. Eleven days is a meaningful shift.
What this means for your work
The most useful framing is to think about the three conditions that make a dynamic workflow worth reaching for: the task is too large for a single context window, you do not know the exact split strategy in advance, and getting the answer right matters more than keeping token costs predictable.
Codebase-wide security audits fit that description well. So do large migrations, framework swaps, API deprecations across thousands of files, and performance optimisation passes driven by profiler output. Klarna noted they have used it to find dead code and surface cleanup opportunities that static analysis missed.
For tasks that are well-defined and repeatable with a predictable token budget, a custom subagent is more efficient. Dynamic workflows are not the right tool for small, bounded jobs where a single agent pass does the job.
One thing to be clear about upfront: dynamic workflows consume meaningfully more usage than a standard Claude Code session. Anthropic recommends starting with a scoped task to get a sense of what that looks like in your specific context before kicking off anything large.
How to start one
There are two ways. You can ask Claude directly, something like “Create a workflow to audit this codebase for unused dependencies”, and Claude will write and run the script. Or you can enable ultracode mode through the effort menu, which sets effort to xhigh and lets Claude decide automatically when a task warrants a workflow. In ultracode mode, Claude plans an understand-change-verify loop on its own: one cluster to map the architecture, a second to execute changes, a third to verify results.
The quickest way to see a workflow in action without any setup is to run /deep-research. That built-in workflow investigates a question across many sources, running agents through defined phases in the background, and delivers a single report at the end.
When a workflow starts for the first time in your session, Claude Code surfaces exactly what is about to run and waits for your confirmation before proceeding. That gate is always there, so you cannot accidentally trigger a 200-agent session from a casual question.
Availability and plan details
Dynamic Workflows are available now in research preview across the Claude Code CLI, Desktop app, and VS Code extension. They are on by default for Max and Team plan users, and for API access via Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Enterprise customers need an administrator to enable the feature through managed settings.
You need Claude Code v2.1.154 or later. If you want to turn workflows off entirely for your organisation, set "disableWorkflows": true in managed settings or use the toggle on the admin settings page.
Anthropic also released Opus 4.8 alongside this launch, an update to Opus 4.7 that arrives just 41 days after the previous model. The new model is reported to be roughly four times less likely to pass flawed code without flagging it, which pairs well with workflows that are specifically designed to run adversarial verification before results land in your hands.
The verification loop is the part worth paying attention to
One detail that gets overlooked in coverage of the sheer scale of these workflows is the adversarial checking built into the design. Agents tackle the problem from independent angles. Other agents then actively try to refute what the first group found. The run keeps iterating until answers converge.
For critical work where a wrong answer has real consequences, that independent verification is the actual value, not just the parallelism. Parallel agents working faster is useful. Parallel agents checking each other’s work before you see any of it is a different kind of useful.
Workflows are in research preview, which means the feature is real and available but Anthropic is still gathering feedback on behaviour at scale. If you are on a qualifying plan and have a task that fits the criteria above, this is worth trying on something scoped before you point it at your most sensitive codebase.