TL;DR
Anthropic says Claude Code can now create dynamic workflows: task-specific JavaScript harnesses that coordinate multiple subagents inside one job. The company presents the feature as useful for complex, high-value work, while warning that it can consume far more tokens than a single-agent task.
Anthropic says Claude Code can now build dynamic workflows, a capability that lets Claude write task-specific orchestration code and coordinate multiple subagents during a single complex job. The feature matters because it moves Claude Code beyond a single agent working through one context window and toward temporary, task-built agent teams for research, review, coding, and other high-value work.
The capability was described by Anthropic’s Claude Code team in a June 2, 2026 Claude blog post titled “A harness for every task: dynamic workflows in Claude Code”, by Thariq Shihipar and Sid Bidasaria. According to the source material, Claude can write a small JavaScript harness that spawns, coordinates, and reviews subagents, using ordinary JavaScript alongside special workflow functions.
In Anthropic’s framing, each subagent can receive a focused brief, operate in a separate context, and in some cases use a different model profile. The workflow can fan out work in parallel, wait at a barrier, then synthesize structured outputs into a final result. The company presents this as a way to reduce familiar single-agent failure modes, including partial completion, self-review bias, and losing track of the original objective over long tasks.
The feature is not being pitched as the default mode for routine prompts. The Thorsten Meyer AI report, citing Anthropic’s caveat, says dynamic workflows use meaningfully more tokens and are intended for complex, high-value tasks, not small edits or simple requests. Confirmed mechanics include workflow patterns such as route-by-task classification, parallel fan-out and synthesis, adversarial checking, generate-and-filter selection, tournament judging, and spawning work until a stop condition is met.
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.
Agent Work Gets More Managed
The main shift is that Claude Code is being positioned less like one worker and more like a system that can assign roles inside a job. For readers who use AI tools for software engineering, research, security review, or operations work, that changes the question from “What prompt should I write?” to “What workflow should manage this task?”
That could matter most for jobs where parallel work and independent review are valuable: large migrations, broad refactors, cited research reports, claim checking, backlog triage, incident post-mortems, security pattern review, naming exercises, and model routing. Anthropic’s reported patterns also suggest a clearer path for separating duties, such as keeping agents that read untrusted public material away from privileged actions.
The tradeoff is cost and control. A dynamic workflow can produce more coverage and stronger review, but it can also spawn many agents and burn far more tokens. The practical value will depend on whether users can set budgets, test workflows on smaller samples, and recognize when a task truly benefits from a coordinated team rather than a single agent.
AI workflow automation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Claude Code’s Third Workflow Axis
The Thorsten Meyer AI report frames dynamic workflows as the third part of a loose trilogy from the Claude Code team. In that framing, skills package organizational knowledge, loops decide how long delegation continues over time, and dynamic workflows handle within-task orchestration.
The background matters because agentic coding tools have often struggled when a job becomes long, parallel, or judgment-heavy. The source material identifies three recurring problems: an agent may declare work done after partial completion, it may favor its own output when asked to review it, or it may lose constraints after long exchanges and context summarization.
Dynamic workflows are presented as a response to those limits. Rather than asking one agent to plan, execute, check, and judge its own work, the system can split the job into roles: an orchestrator, specialists, reviewers, and judges. Those roles are temporary and task-specific, then the team is disbanded when the job is complete.
“A harness for every task: dynamic workflows in Claude Code”
— Anthropic Claude blog post title
multi-agent AI system software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Limits Still Developing
Several practical details remain unclear from the source material. It is not yet clear how widely dynamic workflows are available across Claude Code users, what exact controls administrators have over token budgets, or how teams should audit workflows that Claude writes during a task.
It is also not confirmed how reliably the approach improves outcomes across different domains. Anthropic and the Thorsten Meyer AI report describe patterns and use cases, but readers should treat broad performance gains as claims that need task-specific validation. The strongest near-term evidence will likely come from controlled comparisons on real migrations, reviews, research jobs, and operations queues.
AI orchestration platforms
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Teams Will Test Boundaries
The next step for users is likely careful adoption on bounded tasks: running pilots, setting token limits, comparing single-agent and multi-agent results, and deciding which work justifies the added cost. The Thorsten Meyer AI report points readers to code.claude.com/docs for Claude Code documentation as the feature continues to evolve.
For Anthropic, the larger test is whether Claude Code can make dynamic orchestration predictable enough for production work. If the feature proves reliable, it could become a common pattern for large reviews, research workflows, and complex coding jobs. If costs or coordination errors outweigh the gains, it may remain a specialist tool for high-stakes tasks.
complex AI task management software
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 described dynamic workflows, a Claude Code capability that lets Claude write and run task-specific JavaScript orchestration to coordinate subagents during complex work.
Is this the same as one Claude agent doing a task?
No. The reported feature lets Claude assemble multiple focused subagents, each with its own brief and context, then combine or review their outputs through a workflow.
What kinds of tasks are dynamic workflows meant for?
The source material points to large refactors, deep research, fact-checking, ticket ranking, post-mortems, backlog triage, design review, and security separation patterns. It is aimed at complex work, not small edits.
What is the main downside?
The main downside is higher token use. Anthropic’s caveat, as cited in the report, is that dynamic workflows can cost far more than ordinary single-agent work.
What remains unknown?
It remains unclear how broadly the feature is available, how teams should govern Claude-written workflows, and how much performance improves across real production tasks compared with simpler single-agent runs.
Source: Thorsten Meyer AI
