Dynamic Workflows in Claude Code: When an Agent Swarm Takes Over the Work
Anthropic released dynamic workflows in Claude Code on May 28, 2026. The feature distributes complex tasks across a swarm of parallel AI agents, has them verify each other's findings, and keeps running in the background while the session stays responsive. This article explains how the swarm works, what becomes possible, and what European companies should watch for on cost and governance.
Dynamic workflows in Claude Code, released on May 28, 2026 alongside Claude Opus 4.8, let Claude write a JavaScript orchestration script per task that coordinates tens to hundreds of subagents in parallel. The architecture relies on adversarial verification: one subagent makes a claim, another tries to refute it, and the run iterates until results converge. Hard limits are 16 concurrent and a maximum of 1,000 agents per run. As a real-world example the Bun developer ported the runtime from Zig to Rust: about 750,000 lines of Rust, 99.8 percent of tests passing, 11 days to merge. The central caveat is resource consumption: dynamic workflows use substantially more tokens than a normal session, which is why the feature is off by default for Enterprise plans at launch. For companies the bottleneck shifts from writing the code to confirming that it is correct.
What dynamic workflows in Claude Code are
Dynamic workflows are a feature released on May 28, 2026 that lets Claude Code distribute complex tasks across tens to hundreds of parallel subagents and check their work before a result comes back. Instead of a fixed sequence, Claude writes a JavaScript script per task that runs the orchestration in the background while the session stays responsive. The feature shipped alongside Claude Opus 4.8 and is in research preview.
The feature is activated by a direct request or through the ultracode setting, which handles orchestration automatically. For Max and Team plans it is on by default, for Enterprise it is off at launch and can only be enabled by administrators. This places the release in the trajectory innobu described in the piece on the Agent Control Plane 2026 and the race for the harness .
How the agent swarm works
The core is a self-written script that distributes work and merges results. Claude plans dynamically based on the task, breaks it into subtasks, starts subagents in parallel, and has other subagents refute the findings. The run repeats until the answers stay stable. Three principles shape the behavior.
Parallel and pipeline
Tasks run concurrently or as a pipeline, where each item moves through several stages independently. That keeps total time close to the slowest single chain rather than the sum of all steps.
Adversarial verification
If an agent claims a race condition, another agent's job is to disprove it. That lowers the rate of plausible but wrong findings, because every claim gets an opponent.
Convergence over fixed rounds
The number of agents and iterations is decided at runtime based on the task. The run ends when the answers stop changing, not after a preset number of rounds.
On top of that comes robustness in operation. Progress is saved automatically, interrupted runs resume rather than restart, and coordination happens outside the conversation window so the plan stays stable even on long runs. That explains the image of a developer who kicks off a run and turns to something else.
The hard limits are 16 agents concurrently and a maximum of 1,000 agents per run. The first protects your local machine, the second is a safety net against runaway loops. You should know both numbers before starting a large run, because they set the pace and the maximum effort.
What becomes possible
Dynamic workflows target tasks too large for a single agent: codebase-wide bug hunts, large migrations, security audits, performance reviews, and architecture reviews across thousands of files. The appeal is that a developer can kick off a long run and turn to something else, because verification is built in.
The most prominent example is porting the Bun runtime from the Zig language to Rust. In the run one workflow mapped the Rust lifetimes while a second wrote hundreds of files in parallel, with two reviewers per file. Typical uses according to Anthropic are framework swaps, API deprecations, profiler-guided optimization, and modernizing large legacy systems.
Dynamic workflows fill the gap between firing off a single subagent and building out a complete agent team. Plan to implementation just flows, so we can trust longer runs.
Ken Takao, Lead Systems Engineer, AnthropicThe framing matters: this does not replace every developer, it shifts the role. Whoever used to distribute subtasks and merge results themselves now defines the task and checks the result. How this second generation of agents behaves in practice is covered in the article on the rebuild era of enterprise AI agents .
European perspective
For European companies the central question is not whether the technology impresses, but whether it can be run in a controlled and traceable way. A run with hundreds of subagents produces many automated decisions, and the EU AI Act requires documentation, traceability, and human oversight for high-risk applications. The built-in verification helps but does not replace governance.
Verification becomes the core task: When code appears in minutes, the bottleneck shifts from writing it to confirming that it is correct. With hundreds of parallel findings, evidence capture has to keep pace with generation, or a pile of unverified results builds up.
Data protection follows. If a workflow runs over Amazon Bedrock, Google Vertex AI, or Microsoft Foundry, the same questions about data location and processing apply as with any cloud AI. EU-regional or self-hosted options reduce the risk. Anyone planning a deployment should think it through together with the deadlines of the EU AI Act for high-risk systems .
Enable budget deliberately: That the feature is off by default for Enterprise plans is not a flaw but a safeguard. Administrators can control budget, approval, and data flow before a team starts a run with up to 1,000 agents.
Challenges and risks
The feature is a research preview, not a finished product, and it carries real risks. Four points companies should assess soberly before using dynamic workflows in critical paths.
Token escalation
Dynamic workflows consume substantially more tokens than a normal session, and a run can spawn up to 1,000 agents. Agentic tools can consume many times the tokens of a simple query. Without set budget limits, unexpected bills can result, which is why Anthropic explicitly recommends starting with small, well-scoped tasks.
Verification stays hard
Adversarial checking lowers the error rate but does not guarantee correctness. With hundreds of parallel findings you need your own review and approval processes. The technology moves the hard part from generation to confirmation, and that confirmation cannot be fully automated.
Caution in the enterprise
Reports of pulled-back AI spending, such as the cancellation of many internal Claude Code licenses at Microsoft, show that companies weigh value against cost carefully. A tool that can do a lot but also costs a lot needs a demonstrable business case, not just an impressive demo.
Maturity of the preview
As a research preview, behavior, price, and availability can change. Using it in production for critical paths should factor that in and not assume that today's conditions are permanent.
What companies should do now
The sensible entry point is small and controlled. Anyone who wants to try dynamic workflows should start with a scoped, well-measurable task, set budget limits, and check the results independently before touching production paths. Four steps help.
-
Pilot with a clear boundary
Choose a concrete, measurable task, such as an audit of a scoped module or a limited migration. Define success in advance so you can judge the result clearly later.
-
Control budget and approval
Enable the feature deliberately on Enterprise plans and set token limits and approvals. That keeps a single run with up to 1,000 agents from quietly blowing the budget.
-
Institutionalize verification
Name a human reviewer for the outputs and spot-check findings. The built-in adversarial verification is a help, not a replacement for expert control.
-
Build in governance
Clarify risk classification under the EU AI Act, logging of runs, and data location from the start. That is cheaper when it happens before the first production run, not after.
Anyone planning the step from individual routines to orchestrated runs will find more context in the articles on Claude Code routines and cloud automation and on the pricing shift in AI coding tools and token billing .
Dynamic workflows are not a set-and-forget feature but a tool with high leverage and high cost. The value comes from the right task, clear budget limits, and human review. Get that right and you can speed up large tasks without giving up control.
Further reading
Frequently asked questions
Dynamic workflows are a feature Anthropic released on May 28, 2026, that lets Claude Code distribute complex tasks across tens to hundreds of parallel subagents and check their work before returning a result. Instead of a fixed sequence, Claude writes a JavaScript script per task that runs the orchestration in the background while the session stays responsive. The feature shipped alongside Claude Opus 4.8 and is in research preview.
A maximum of 16 agents concurrently and a maximum of 1,000 agents per run. The first limit protects local resources, the second prevents runaway loops. In practice a workflow coordinates tens to hundreds of subagents, whose count and iterations are decided at runtime based on what the task needs. The run repeats until the results converge.
As a real-world example Anthropic cites porting the Bun runtime from the Zig language to Rust. The result was about 750,000 lines of Rust, 99.8 percent of the existing tests passing, and 11 days from first commit to merge. In the example one workflow mapped Rust lifetimes while a second wrote hundreds of files in parallel, with two reviewers per file.
Dynamic workflows consume substantially more tokens than a normal Claude Code session because a run can spawn up to 1,000 agents. Anthropic explicitly recommends starting with small, well-scoped tasks. For Enterprise plans the feature is off by default at launch so administrators can control budget and approval. Without budget limits, unexpected bills can result.
For clearly scoped tasks such as audits, migrations, or bug hunts, yes, but in a controlled way. A small pilot with defined success criteria, set budget limits, and a human reviewer for the results is sensible. For high-risk applications the EU AI Act requires documentation, traceability, and human oversight. The built-in verification helps but does not replace governance.