Anthropic shipped a capability alongside Claude Opus 4.8 that may change how large software teams think about what a coding agent can actually do. Dynamic Workflows, now in research preview inside Claude Code, lets a single session author a JavaScript orchestration script that spawns, directs, and reconciles the output of up to 1,000 parallel subagents. The result is a tool that can take on problems that were previously too large to hand to any single AI session.
Anthropic released the feature on May 28, 2026, requiring Claude Code version 2.1.154 or later. It is available on Max, Team, and Enterprise plans, with Enterprise customers able to opt in through admin controls. The CLI, Desktop app, and VS Code extension all support it from launch.
How Dynamic Workflows Differ from Existing Subagents
Claude Code has long been able to spawn subagents. The existing model works turn by turn: Claude decides which worker to send a task to, waits for a result, reads that result into its own context window, and then decides what to do next. That architecture works well for moderately complex projects, but it has a ceiling. Every intermediate result costs context tokens, and the coordination overhead grows quickly as the number of parallel tasks increases.
Dynamic Workflows move orchestration logic out of Claude's context and into code. When a user starts a workflow, Claude writes a JavaScript script that encodes all the sequencing, branching, and conditional logic the task requires. That script then executes outside the conversation, spinning up agents, collecting their outputs into variables, and resuming from saved checkpoints if a run is interrupted. Claude does not need to hold the state of a thousand concurrent tasks in its own memory, because the script does it instead.
Dynamic Workflows: Key Numbers
- Maximum agents per run1,000
- Maximum concurrent agents16
- Minimum Claude Code versionv2.1.154
- Available plansMax, Team, Enterprise
- Bun Zig-to-Rust port: lines of code750,000
- Bun port: test suite pass rate99.8%
The Bun Port That Put the Feature to the Test
Before the public launch, Jarred Sumner, the creator of Bun, used Dynamic Workflows to port his JavaScript runtime from Zig to Rust. The task was large by any measure: the final output was roughly 750,000 lines of Rust code, and the migration had to preserve compatibility with Bun's existing test suite. Sumner completed the work in 11 days from first commit to merge, with 99.8% of the existing tests passing at the end.
The example is striking not just for its scale but for what it says about the kind of work the feature is suited for. Porting a systems-level runtime between two low-level languages is exactly the sort of deeply interdependent, repository-spanning task that collapses under the token budget of any single conversation. Breaking it into subtasks, each handled by a focused subagent, with the orchestration script keeping track of which pieces are done and which depend on what, is the only way to get through it at this speed.
"A workflow can consume substantially more tokens than a standard Claude Code session because every agent pays its own context overhead. We recommend starting with a well-scoped task to calibrate consumption." Anthropic, Claude Code Dynamic Workflows documentation, May 2026
Practical Limits and Token Considerations
The cap of 16 concurrent agents at any one time is a guardrail against runaway token spend, not a hard architectural limit. Anthropic is explicit that Dynamic Workflows can consume tokens at a rate that surprises users who are accustomed to single-session work. A run that spawns a hundred agents, each working on a moderately complex subtask, will burn through context budgets faster than almost any other Claude Code pattern.
The recommended approach is to begin with a tightly scoped task, observe how many agents the workflow spawns and at what cost, and then scale up. The feature also supports resuming from saved checkpoints, which means an interrupted workflow does not need to restart from scratch. That matters for tasks that run for hours or days and might hit resource limits partway through.
Users can trigger the feature simply by including the word "workflow" in a prompt. Anthropic's documentation also provides explicit syntax for passing orchestration parameters directly. The underlying Claude Code architecture handles the scaffolding, so the user-facing interface is intentionally minimal.
What This Means for Agentic Coding
Dynamic Workflows represent a qualitative shift in the kinds of problems that fit inside a single Claude Code engagement. Prior to this feature, tasks that touched dozens of files across a large monorepo, or that required a consistent set of refactors applied simultaneously across many modules, were difficult to handle without breaking them into manual phases. With an orchestration layer that can run a thousand targeted subagents in parallel, those tasks become more tractable.
The feature also has implications for how teams structure their use of Claude's agentic tools more broadly. A workflow that runs autonomously for eleven days, producing three-quarters of a million lines of working code, is not a prompt. It is closer to a long-running engineering process. Anthropic is careful to note that the feature is still in research preview, and the token economics are a real constraint. But the direction is clear: the ceiling for what a single Claude Code session can accomplish just moved up significantly.
