TL;DR
Anthropic’s Claude Code team has described dynamic workflows, a feature that lets Claude write a task-specific JavaScript harness to coordinate temporary subagents. The approach is aimed at complex, high-value work, not routine requests, and Anthropic says it can use far more tokens.
Anthropic says Claude Code can now create dynamic workflows, allowing Claude to write a task-specific JavaScript orchestration harness that spawns and coordinates temporary subagents for complex work. The development matters because it moves Claude Code beyond a single-agent setup toward multi-agent task execution, though Anthropic says the approach uses meaningfully more tokens and is intended for high-value tasks.
The feature, described by Anthropic in a June 2, 2026 Claude blog post by Thariq Shihipar and Sid Bidasaria, lets Claude build a custom workflow around the task at hand. According to the source material, Claude can create an orchestration layer that routes work, fans tasks out to subagents, waits for results, and then synthesizes structured outputs into a final answer.
Each subagent can receive its own focused brief and separate context window. The source material says this separation is meant to reduce common single-agent failures such as stopping early, judging its own output too favorably, or losing the original goal across long tasks. Anthropic’s mechanics describe subagents that can be used for specialist work, independent review, adversarial checks, or judging competing answers.
The article frames dynamic workflows as the third piece in a broader Claude Code pattern: skills package organizational knowledge, loops decide how far to delegate over time, and dynamic workflows let Claude build a temporary team within a single task. The source material stresses a boundary: this is not meant for simple jobs such as fixing a typo, because the token cost can be much higher than a direct single-agent request.
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.
Claude Code Moves Toward Teams
The main change is architectural: Claude Code is being presented less as one worker handling a task alone and more as a system that can commission multiple agents for different parts of a job. For readers using coding agents or AI assistants at work, that could affect how they plan large refactors, research projects, audits, and triage tasks.
The source material says dynamic workflows may be useful when work is large, parallel, adversarial, or judgment-heavy. Examples given include big migrations, cited research reports, claim-by-claim fact checks, ranking large ticket backlogs, post-mortems, design or naming reviews, and security pattern checks. Those are tasks where a single context window can become overloaded or where independent review may improve quality.
The tradeoff is cost and control. Anthropic’s caveat, as described in the source material, is that workflows can consume far more tokens and may spawn many agents if left unbounded. That means teams adopting the approach will need budgets, stop conditions, and pilot runs before using it for production-scale work.
AI multi-agent workflow automation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Single Agent To Harness
Traditional agent use often places planning, execution, and review inside one long-running context. The source material identifies three failure modes in that pattern: agentic laziness, where an agent declares work finished too early; self-preferential bias, where it favors its own output; and goal drift, where original constraints fade during long sessions or after summarization.
Dynamic workflows address those problems by separating roles. The workflow can act like a dispatcher, assign subagents to narrow tasks, add independent reviewers, and merge results only after a barrier step where all work has completed. Anthropic’s mechanics include patterns such as classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournaments, and loop-until-done execution.
The source material also highlights a security pattern: quarantine. Agents that read untrusted public content should be barred from high-privilege actions, while a separate agent performs those actions. That framing is defensive and centers on separation of duties for autonomous agent systems.
“Claude writes its own harness and assembles a temporary team of subagents.”
— Thorsten Meyer AI, summarizing Anthropic’s Claude Code post
JavaScript orchestration tools for AI workflows
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Reliability Still Open
Several details remain unclear from the supplied material. It does not give benchmark results, average token costs, or measured reliability gains for dynamic workflows compared with a single Claude Code agent. The available description explains the mechanism and intended use cases, but does not quantify how often the approach improves outcomes.
It is also unclear how teams should set default limits for agent counts, token budgets, model routing, and stop conditions. The source material warns that workflows can spawn large numbers of agents, so operational guardrails may be a central factor in whether the feature is useful or too expensive for routine work.
AI subagent coordination platform
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Pilots Before Broad Use
The next step for Claude Code users is likely small-scale testing on tasks where multi-agent work has a clear reason: large refactors, independent verification, high-volume triage, or research that needs structured review. The source material recommends bounding workflows with token budgets and pilots before broader use.
Future updates to Anthropic’s Claude Code documentation may clarify recommended patterns, cost controls, and safety practices. For now, the confirmed takeaway is narrow: dynamic workflows are available as a recent, still-developing Claude Code capability, aimed at complex tasks where a single agent is more likely to stop short, grade its own work, or lose the plot.
complex AI task management tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Anthropic announce?
Anthropic described dynamic workflows in Claude Code, a capability that lets Claude write a task-specific JavaScript harness to spawn, coordinate, and synthesize work from subagents.
Is this meant for everyday Claude tasks?
No. The source material says dynamic workflows are built for complex, high-value tasks and can use meaningfully more tokens. Simple requests are better handled by a direct single-agent approach.
What kinds of work could benefit?
The supplied examples include large migrations, deep research with citations, claim checking, backlog ranking, root-cause reviews, design evaluation, and security-oriented review patterns.
What is confirmed versus claimed?
Confirmed from the source material: Anthropic has described the feature, its mechanics, and its intended patterns. Claimed or still developing: the scale of quality gains, real-world cost ranges, and best default limits for production teams.
What should teams watch next?
Teams should watch for updated Claude Code documentation, clearer budget guidance, and field reports showing when dynamic workflows outperform a single agent enough to justify the added cost.
Source: Thorsten Meyer AI
