TL;DR
Anthropic’s Claude Code team has described a feature called dynamic workflows, in which Claude can write a task-specific JavaScript harness and coordinate temporary subagents. The approach is aimed at complex, high-value work, but Anthropic says it uses meaningfully more tokens and is not meant for simple tasks.
Anthropic’s Claude Code team has detailed a feature called dynamic workflows that lets Claude write its own orchestration harness and coordinate temporary subagents for complex tasks, a development that matters because it moves some AI work from a single-agent model toward task-specific agent teams.
The feature was described by Anthropic in a Claude blog post titled “A harness for every task: dynamic workflows in Claude Code,” published on June 2, 2026 by Thariq Shihipar and Sid Bidasaria. According to the source material, a dynamic workflow is a small JavaScript program that Claude writes and runs to spawn, coordinate, and merge the work of subagents.
The reported goal is to reduce common single-agent failure modes on larger tasks, including partial completion, self-review bias, and goal drift across long runs. Each subagent can receive a focused brief and a separate context window, while another agent or workflow step can review, compare, or synthesize the results.
Anthropic’s own caveat is central to the announcement: dynamic workflows use meaningfully more tokens and are aimed at complex, high-value tasks, not routine edits or simple requests. The source material frames the feature as part of a wider Claude Code pattern alongside skills, which package organizational knowledge, and loops, which decide how long delegation should continue over time.
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 Toward Agent Teams
The development matters because it changes the operating model for some AI-assisted work. Instead of asking one agent to plan, execute, verify, and revise inside a single context, Claude Code can create a temporary division of labor for one task and shut it down when the work is complete.
That could be useful for work that is large, parallel, adversarial, or judgment-heavy. The source material lists examples including big migrations and refactors, deep research reports, claim-by-claim fact-checking, ranking large ticket backlogs, root-cause post-mortems, security reviews, and model routing.
The practical tradeoff is cost and control. More subagents can mean more coverage and independent review, but also more token use and more need for boundaries such as budgets, pilot runs, stop conditions, and explicit review stages. For readers using AI tools at work, the news is less about replacing prompts with bigger prompts and more about deciding when a task needs structured coordination.
AI Tools for Everyday Tasks: The Complete Beginner’s Guide To Working Smarter with AI
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Claude Coordinates Subagents
According to the source material, dynamic workflows can combine several orchestration patterns. These include classify-and-act, where work is routed by task type; fan-out-and-synthesize, where parallel agents work independently before a merge step; and adversarial verification, where a separate agent is assigned to challenge a result.
Other patterns include generate-and-filter, in which many ideas are produced and narrowed; tournament, where agents compete and are judged in pairs; and loop-until-done, where agents continue until a stop condition is met rather than a fixed number of iterations.
The source material also highlights a security pattern described as quarantine: agents that read untrusted public content should be separated from agents allowed to take high-privilege actions. That separation of duties is presented as a way to reduce risk when autonomous systems handle outside information and sensitive operations in the same broader workflow.
The Claude Code Sub-Agent Playbook: 100+ Battle-Tested Workflows to Ship Faster, Cut Token Burn, and Automate Real Engineering
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits Still Need Testing
Several details remain unclear from the provided source material. It is not yet clear how broadly dynamic workflows are available across Claude Code users, what default limits Anthropic applies, or how reliably Claude chooses the right workflow shape without human steering.
It is also unclear how the approach performs across different domains, especially where errors are costly. The claims about reducing agentic laziness, self-review bias, and goal drift are presented as the rationale for the feature, but the source material does not provide independent benchmark results or failure-rate comparisons.
Intelligent Agentic AI Systems with JavaScript in 2026: Designing Autonomous Workflows with LLMs, Tools, and Real-World Integrations (Practical AI Systems Engineering Series Book 2)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Users Will Test Boundaries
The next milestone is practical adoption inside Claude Code workflows. Teams using the feature will need to decide when a task is worth the added cost, how to set token budgets, and when to require independent review before acting on outputs.
Readers should watch for Anthropic documentation updates, real-world examples from developers, and evidence on whether dynamic workflows improve outcomes enough to justify their added complexity. For now, the confirmed development is that Claude Code can create task-specific subagent workflows; the open question is how often that structure outperforms a simpler single-agent run.
The LangGraph Mastery Guide: Orchestrating Multi-Agent AI Workflows for Complex Task Automation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are Claude dynamic workflows?
Dynamic workflows are task-specific orchestration programs that Claude can write and run in Claude Code to coordinate multiple subagents during a complex job.
Is this meant for everyday Claude tasks?
No. According to the source material, Anthropic says the feature uses meaningfully more tokens and is intended for complex, high-value tasks, not simple edits or routine prompts.
What problems is Anthropic trying to address?
The feature is aimed at reducing problems seen in long single-agent runs, including partial completion, self-review bias, and goal drift after extended work or context compression.
What kinds of work could use dynamic workflows?
The source material names large refactors, deep research, fact-checking, backlog triage, security review, root-cause analysis, and judgment-heavy tasks where parallel work and independent review may help.
What remains uncertain?
It remains unclear how broadly the feature is available, how much it costs in typical use, and how reliably it improves results compared with a single-agent workflow across different tasks.
Source: Thorsten Meyer AI
