Claude Code Dynamic Workflows: the agent swarm enters research preview
On 28 May 2026, Anthropic opened the research preview of “Dynamic Workflows” in Claude Code — alongside the release of Opus 4.8. A dynamic workflow is not a new model but an orchestration mechanism: at runtime, Claude writes a JavaScript script for the task you describe, and a separate runtime executes that script in the background while your session stays responsive. Instead of working through a complex task turn by turn inside a single context window, the script fans the work out across many subagents that run in parallel and review each other. The interesting question is not whether the model got stronger — it is whether the architecture on top is what decides the outcome.
What actually changed
- A dynamic workflow is a JavaScript script that Claude writes itself at runtime. An isolated runtime executes it in the background; intermediate results live in script variables rather than the context window.
- It scales to up to 1,000 subagents per run, at most 16 concurrently. The first cap prevents runaway loops, the second bounds local CPU load.
- It is triggered by the word “workflow” in your prompt or the
/effort ultracodecommand — the latter plans a workflow automatically for every substantive task. - It requires Claude Code v2.1.154 or later. Available in research preview on all paid plans, including Pro (enable it there via
/config) — among them Max, Team, and Enterprise — plus the API and the partner platforms Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. - Independent agents review each other adversarially: one agent tries to refute what another found before the result is reported back.
What applied before
Until now, Claude Code coordinated multi-step tasks through subagents, skills, and agent teams — but in all three cases Claude itself remained the orchestrator. The model decided turn by turn which worker to spawn next, and every intermediate result landed in the context window of the running conversation. That works for a handful of delegated tasks per turn, but it hits a wall once a task needs more agents than one conversation can cleanly coordinate: the context window fills up with intermediate state, and the orchestration itself is not repeatable — it lives in the conversation transcript, not in an artifact.
So anyone who wanted to run a codebase-wide bug sweep or a migration across hundreds of files had to either break up the parallelism by hand or build their own orchestration scaffold.
What applies now
With the research preview, Anthropic moves the plan out of the context window and into code:
1. The plan lives in the script, not the context. For the task you describe, Claude writes a JavaScript script that holds the loop, the branching, and the intermediate results itself. The runtime executes it in isolation from the conversation. Claude’s context window only ever sees the final answer — not the transcript of hundreds of agents. The script is written under the session directory, so it is readable, versionable, and rerunnable.
2. Scale with hard limits. A single run may spawn up to 1,000 subagents in total, but at most 16 concurrently (fewer on machines with few CPU cores). The 1,000 cap is explicitly documented as a guard against runaway loops; the 16-way concurrency is there to protect local resources.
3. Quality through confrontation, not more passes. Because the plan lives in code, a workflow can apply a repeatable quality pattern: independent agents review each other’s findings adversarially before anything is reported; a hard plan is drafted from several angles and weighed against each other. According to Anthropic, that is what separates this from “just running more agents.”
4. Triggers and prerequisites. A workflow is triggered by the word “workflow” in your prompt, or by /effort ultracode, which combines xhigh reasoning with automatic orchestration. It requires Claude Code v2.1.154+. Anthropic names widely scattered bugs, large migrations, security audits, performance reviews, and architecture analysis as the use cases.
Why it matters
The truly notable thing about Dynamic Workflows is not the number 1,000. It is that Anthropic deliberately positions a feature shipping alongside its new flagship model Opus 4.8 as an architecture tool — not a model upgrade. That fits a thesis that has been circulating in the trade press for months: AI agents rarely fail because of the model, but because of their architecture. t3n lists “model focus instead of context engineering” as one of its five core failures — what matters is not which model you pick, but which data the agent sees at each step. By that logic, an agent on the weaker model with excellent context engineering beats one on the strongest model with poor context.
Context engineering means planning the information for each step architecturally — precise task prompts, a retrieval system that pulls in the right context, and an intermediate evaluation after each step. Those are exactly the building blocks inside Dynamic Workflows: splitting work into sharply scoped subtasks, pulling sources in parallel (the bundled /deep-research workflow fetches and cross-checks sources against each other), and adversarial review before reporting back. Anyone who wants to understand why this is more than “more compute” quickly arrives at the fundamentals of context engineering and agentic RAG — retrieval that does not run once and blindly, but iterates under the agent’s control. In practice, the quality of such pipelines often hinges on details like reranking and contextual retrieval, which decide whether the right evidence reaches the agent at all.
The sober reading: with Dynamic Workflows, Anthropic shifts the lever from “better model” to “better orchestration.” That lines up with the model-versus-agent distinction it-daily draws as well: models recognise patterns, agents orchestrate workflows and reach for models along the way. Take that seriously and you tinker less with the prompt and more with the system — with decomposition, retrieval, and review loops. At boostN we work on exactly these architecture and workflow patterns; in that sense the release confirms a direction we consider more durable than waiting for the next model. What remains open is whether adversarial self-review actually proves more reliable in practice or just more expensive — the token cost of a run with hundreds of agents is substantial, and Anthropic deliberately asks for confirmation before the first run.
What you can do now
If you want to test Dynamic Workflows: Update Claude Code to v2.1.154+ and start with the bundled /deep-research workflow or the word “workflow” in your prompt on a small slice — a single directory rather than the whole repo. The /workflows view shows each agent’s token usage live; stop the run there at any time without losing completed work.
If the cost gives you pause: Treat a workflow like an investment, not a standard request. A run with many agents uses substantially more tokens than the same task in conversation and counts toward your plan’s limits. Test on a narrow slice before launching a large migration.
If the architecture thesis interests you: Look at how an agent system is sensibly built — we cover the building blocks (task decomposition, retrieval, evaluation) in detail in the lexikon article on building AI agents.
Fundamentals in the lexikon
How retrieval-augmented generation works and why it is the core of good agent architecture is covered here: → Understanding RAG — retrieval-augmented generation
Entdecke mehr
Why AI Models Find Different Code Problems
Three frontier models, the same 1000-line script, three different finding lists — and why that very spread makes multi-orchestration strong.
GlossarEnsemble / Multi-Model Orchestration
Ensemble means combining several deliberately varied LLM runs or models whose findings complement each other. Multi-model orchestration drives these runs via orchestrators with sub-agents, so the union of results is larger than any single run.
LexikonFunction Calling / Tool Use
How an LLM uses tools: define a tool as a schema, the model picks the function and arguments, the result returns to the chat — the basis of every agent.