TL;DR
Anthropic has described dynamic workflows in Claude Code, a capability that lets Claude write and run a task-specific JavaScript harness to coordinate multiple subagents. The company presents it as a fit for complex, high-value tasks, while warning that it can use far more tokens than a single-agent run.
Anthropic’s Claude Code team has detailed a feature called dynamic workflows that lets Claude write its own task-specific orchestration code and coordinate multiple subagents during a single job, a development aimed at complex work where one long-running agent may lose accuracy, miss steps, or grade its own output too favorably.
The system, described by Anthropic in a June 2 Claude blog post by Thariq Shihipar and Sid Bidasaria, works by having Claude generate a small JavaScript harness for the task at hand. That harness can spawn specialized agents, give each one a focused brief, wait for their outputs, and combine the results into a final answer or action plan.
Thorsten Meyer AI described the feature as Claude creating an org chart for one job: a dispatcher, specialists, reviewers, judges, or other temporary roles that exist only until the work is finished. The mechanics and patterns are attributed to Anthropic; that org-chart framing is the site’s interpretation of how readers should think about the system.
Anthropic’s stated caveat is that dynamic workflows use meaningfully more tokens and are intended for complex, high-value tasks. The source material explicitly warns that this is not the right tool for small edits, such as asking Claude to fix a typo.
When one agent isn’t enough: Claude now builds its own team on the fly
Skills package what you know; loops decide how far you delegate over time. Dynamic workflows are the third axis — within a single task, Claude writes its own harness and assembles a temporary team of subagents. Think of it as Claude drawing an org chart for one job.
The shift is from prompting a worker to commissioning a team — more output, more cost, and a manager’s judgment required. Reach for a workflow when a task is big, parallel, adversarial, or judgment-heavy — and when you can feel a single agent getting lazy, grading its own homework, or losing the plot. Bound it (token budgets, pilot first) — workflows can spawn hundreds of agents and burn far more tokens. For everything else, don’t hire five people to change a lightbulb.
A Shift From Worker To Team
The development matters because it changes the role of the coding agent from a single assistant completing a task inside one context window to a manager of multiple agents. For difficult jobs, that can help separate work into smaller pieces, reduce missed steps, and add independent review before the final result is returned.
The source material identifies three recurring problems with single-agent work on large tasks: agentic laziness, where work stops before all items are finished; self-preferential bias, where an agent evaluates its own work too positively; and goal drift, where the original objective weakens across long runs or summarized context. Dynamic workflows are presented as a way to reduce those risks by assigning separate agents isolated briefs and review duties.
The cost side is also part of the news. A workflow that fans out to many agents can consume far more tokens than a single-agent session. That means users and organizations may need clearer judgment about when the extra coordination is justified.
Advanced Automation with JavaScript: Techniques for Modern Task Management alongside In-Depth Strategies for Automating and Optimizing Your Workflows
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Claude Code Fits In
The feature sits in what Thorsten Meyer AI called a loose trilogy from Anthropic’s Claude Code team. In that framing, skills package an organization’s knowledge, loops decide how far to delegate over time, and dynamic workflows handle the structure of delegation inside one task.
Anthropic’s workflow patterns include classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop-until-done. In practical terms, the system can route work by task type, run agents in parallel, assign a separate reviewer to attack results, or keep spawning work until a stop condition is met.
The cited use cases go beyond coding. The source material lists large migrations, deep research reports, claim-by-claim fact-checking, ticket ranking, root-cause post-mortems, backlog triage, naming or design by rubric, model routing, and security review patterns as areas where dynamic workflows may earn their cost.
“Claude writes its own harness and assembles a temporary team of subagents.”
— Thorsten Meyer AI
AI workflow orchestration software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Limits Still Matter
It is not yet clear from the supplied material how broadly dynamic workflows are available across Claude Code users, how the feature behaves across all task types, or how often it improves results enough to justify the added token cost. The source presents the feature as recent and still developing.
Performance claims should also be read as task-dependent. Anthropic and the source material describe patterns and likely use cases, but they do not establish that every multi-agent workflow will outperform a well-scoped single-agent prompt. The benefit appears strongest when work is parallel, adversarial, judgment-heavy, or long enough for drift to become a risk.
Security boundaries remain a live concern. The source material highlights a quarantine pattern: agents that read untrusted public content should be barred from high-privilege actions, while a separate agent performs the action. That is a design recommendation, not proof that every implementation will be safe by default.
multi-agent AI system
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Teams Will Need Guardrails
The next step for users is likely experimentation with bounded pilots: small workflow trials, token budgets, clear stop conditions, and independent checks on whether the multi-agent structure improves the final result. Teams using Claude Code will also need to decide which tasks deserve this mode and which should remain simple single-agent jobs.
Anthropic’s documentation at code.claude.com/docs is the place to watch for implementation details, availability notes, and workflow guidance. For now, the practical takeaway is narrow: dynamic workflows appear designed for jobs where a single agent is likely to miss work, lose the thread, or need an independent reviewer.
AI task coordination tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Anthropic announce about Claude Code?
Anthropic detailed dynamic workflows, a Claude Code capability that lets Claude write a task-specific JavaScript harness and coordinate multiple subagents during one complex task.
Does this mean Claude always uses multiple agents now?
No. The source material says dynamic workflows are meant for complex, high-value work. Routine tasks, such as small edits, are still better suited to a simpler single-agent approach.
Why would multiple subagents help?
Separate agents can work in parallel, focus on narrower briefs, and review one another’s output. That may reduce risks such as partial completion, self-review bias, and goal drift during long tasks.
What is the main downside?
The main tradeoff is cost. Anthropic’s caveat, as cited in the source material, is that workflows can use meaningfully more tokens than a single-agent run.
What kinds of work are best suited to dynamic workflows?
The cited examples include large refactors, research reports, fact-checking, backlog triage, root-cause analysis, security review, model routing, and other tasks that are large, parallel, adversarial, or judgment-heavy.
Source: Thorsten Meyer AI
