Anthropic shipped Claude Code Dynamic Workflows as a research preview on May 28, 2026, and the feature is architecturally more consequential than the Opus 4.8 benchmarks that dominated most coverage that day. The reason: Dynamic Workflows does not just add more agents to an existing system. It moves the orchestration plan out of the model's context window and into a JavaScript script that Claude writes on the fly — a distinction that determines whether a 500-agent run is even feasible, and whether your next invoice is predictable.

For a technically literate developer on Max, Team, or Enterprise, this changes how Claude Code behaves at scale. For a developer on Pro, it requires a deliberate step to turn on. For everyone, it introduces a cost dynamic that Anthropic itself flags explicitly: Dynamic Workflows consumes substantially more tokens than a normal Claude Code session, and a large run at Opus 4.8 pricing can change the bill by an order of magnitude.

Plan Lives in Code, Not Context Window

The core architectural shift in Dynamic Workflows is a single design decision: the orchestration plan moves into executable code rather than living in the model's working memory. In previous Claude Code patterns, subagents were dispatched by Claude turn by turn, with every intermediate result accumulating in the context window. That context accumulation was the binding constraint on autonomous long-running tasks — it capped how many agents could coordinate before the window filled.

In Dynamic Workflows, Claude writes a JavaScript orchestration script for the task you describe. A background runtime then executes that script. The orchestration logic — loops, branching, agent-count decisions, verification passes — lives in script variables. The model's context window receives only the final, verified answer. That separation is what makes runs involving tens to hundreds of parallel subagents viable without exhausting the context ceiling that previously made large autonomous tasks impractical.

The security model follows from this structure. The workflow script itself has no direct access to the filesystem or the shell — only the subagents read, write, and execute commands, and those subagents always operate in acceptEdits mode, inheriting the session's existing tool permissions. Progress is saved as execution proceeds, so an interrupted run picks up where it left off rather than starting over.

Claude Code Dynamic Workflows: Execution Limits and How Agents Reach an Answer

Anthropic's documentation sets explicit caps on concurrent agent usage. According to the official Claude Code documentation, a workflow can run up to 16 agents concurrently, with a hard ceiling of 1,000 total agents per run. The official launch post describes runs spanning "tens to hundreds of parallel subagents," consistent with those limits.

The verification architecture is the part that most coverage has underweighted. Dynamic Workflows is not a simple fan-out-and-collect system. Once Claude breaks a task into subtasks and dispatches agents to work in parallel — each approaching the problem from an independent angle — a separate layer of agents runs to refute those findings. The run iterates until the results converge under that adversarial check before anything reaches the user. The bundled /deep-research command illustrates the pattern directly: it fans out web searches, cross-checks each source against the others, and applies a survival vote on each claim before producing a cited report.

How Does Ultracode Differ From Triggering a Workflow Manually?

There are two ways to start a Dynamic Workflow. The simpler is to include the word "workflow" in a prompt — Claude Code highlights it and plans a fan-out for the task rather than working through it turn by turn. The more significant is ultracode, a session-level setting accessible through the effort menu.

ultracode combines xhigh reasoning effort with automatic workflow orchestration: it sets Claude to its highest standard effort level and then lets the model decide on its own whether a given task warrants spinning up workflow machinery. With a plain workflow trigger, the developer decides to fan out. With ultracode active, Claude makes that call — running a pre-assessment of each request, then choosing whether to handle it inline or deploy dozens or hundreds of agents to address it from multiple angles simultaneously.

That autonomy is the feature's real editorial significance. The decision to reach for the power tool moves from the human to the model. The official documentation notes that with ultracode on, a single request can produce multiple workflows in sequence: one to map the architecture, a second to make the changes, a third to verify.

Bun Port Demonstrates Scale, Remains Out of Production

The demonstration Anthropic has centered its launch coverage on is real and instructive at scale. Bun creator Jarred Sumner used Dynamic Workflows to port Bun from Zig to Rust: approximately 750,000 lines of Rust, 99.8% of the existing test suite passing, and eleven days from first commit to merge. The workflow mapped correct Rust lifetimes for every struct field in the Zig codebase, then ran hundreds of parallel agents with two independent reviewers per file while a fix loop drove the build and test suite until both passed cleanly.

The caveat belongs in the same sentence: Anthropic's official blog states explicitly that the Bun port is not yet in production. It is a credible proof of concept at a scale no single-agent loop could reach, not a shipped result.

The enterprise testimonials in Anthropic's launch materials offer more immediately applicable context. Alessio Vallero, Senior Engineering Manager at Klarna, described Dynamic Workflows as especially valuable for discovery tasks across large codebases — specifically citing dead-code identification and cleanup opportunities that traditional static analysis missed. Ken Takao, Lead Systems Engineer at CyberAgent, noted the feature fills the gap between firing off a single subagent and building out a full agent team.

Plan Access and Pricing Before Running a Large Workflow

Dynamic Workflows is available across all paid Claude Code plans as of May 28, 2026. The official Claude Code documentation confirms the feature is on by default for Max and Team plans and for users on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Enterprise plans default to off, with administrators able to enable it through managed settings. Pro plan users can turn it on manually from the Dynamic Workflows row in /config. Claude Code version 2.1.154 or later is required on all platforms.

The cost structure deserves deliberate attention before triggering any large run. Each agent in a workflow uses the model active in the current session. If that model is Opus 4.8 — at the same rate Anthropic kept flat from Opus 4.7 — a 500-agent audit can shift the session bill by an order of magnitude compared to a standard Claude Code session. Anthropic recommends checking /model before a large run, setting explicit token budgets, and starting on a scoped task to calibrate consumption before pointing the machinery at an entire codebase. The confirmation dialog that appears when a workflow first triggers — showing the planned phases and agent counts before any execution begins — is an intended safety valve, not a procedural formality.

Early testers have flagged the burn rate explicitly. One developer quoted in post-launch coverage described a session where Claude attempted to launch 47 concurrent agents, successfully launched 25, and made errors over a five-hour run that required human review to catch. The takeaway from that account is not that the feature is broken, but that large-scale parallel orchestration at research-preview maturity still requires active supervision.

Dynamic Workflows is the most architecturally significant thing in the Opus 4.8 release package, because it moves orchestration from something engineers hand-build into something the model writes and runs itself — turning Claude Code from a single coding assistant into a temporary engineering team you direct. It is also a research preview, billed by the token, with its flagship demo still out of production. The capability is real; the discipline of scoping, budgeting, and verifying it is now the job.

Frequently Asked Questions

How do Claude Code Dynamic Workflows work?

When a Dynamic Workflow triggers, Claude writes a JavaScript orchestration script for the task on the fly. A background runtime executes the script while the session stays responsive. Agents work in parallel from independent angles, a separate set of agents attempts to refute their findings, and the run continues iterating until results converge — then delivers a single verified answer to the session. Because the orchestration logic lives in script variables rather than the model's context window, runs can coordinate tens to hundreds of agents without exhausting the context ceiling that limits turn-by-turn agent patterns.

What is ultracode in Claude Code?

ultracode is a Claude Code session setting available through the effort menu. Activating it sets the model's reasoning effort to xhigh and grants Claude the autonomy to decide — for each request in the session — whether the task warrants spinning up a Dynamic Workflow. With a plain workflow trigger, the developer decides when to fan out; with ultracode on, Claude makes that decision automatically, and a single request can produce multiple workflows in sequence covering understanding, execution, and verification phases.

Which plans support Claude Code Dynamic Workflows?

Dynamic Workflows is available on all paid Claude Code plans and through the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. On Max and Team plans it is on by default. Enterprise defaults to off, with administrators enabling it through managed settings. Pro plan users can activate it manually from the Dynamic Workflows row in /config. Claude Code version 2.1.154 or later is required on all platforms.

How much do Dynamic Workflows cost in tokens?

Anthropic has not published a fixed token rate for Dynamic Workflow runs because cost depends on the task scope, agent count, and the model active in the session. Each agent in a workflow pays its own context overhead against the session model's standard rate. A 500-agent run on Opus 4.8 can shift the total bill by an order of magnitude compared to a normal Claude Code session. Anthropic recommends starting with a scoped task to calibrate usage before running workflows against a full codebase.