TL;DR
Anthropic has described a Claude Code capability called dynamic workflows, in which Claude writes a task-specific JavaScript harness to coordinate multiple subagents. The company frames the feature as a tool for complex, high-value work, while warning that it can use far more tokens than a single-agent request.
Anthropic’s Claude Code team has detailed a feature called dynamic workflows that lets Claude write its own orchestration harness for a task, then coordinate multiple subagents to complete and check the work. The development matters because it points to a shift from using one AI agent for long, complex jobs toward temporary, task-specific agent teams.
The feature was described by Anthropic in a June 2, 2026 Claude blog post titled “A harness for every task: dynamic workflows in Claude Code,” according to the source material. A July 1, 2026 article from Thorsten Meyer AI frames the capability as the third part of a broader Claude Code arc: skills package organizational knowledge, loops manage delegation over time, and dynamic workflows coordinate several agents inside a single task.
Mechanically, the source says a dynamic workflow is a small JavaScript program that Claude writes and runs for the job at hand. That program can spawn subagents, assign focused briefs, wait for outputs, merge results and send work to independent checking agents. Each subagent can have its own clean context window, focused goal and, in some cases, a different model choice.
Anthropic’s stated caveat is material: this approach uses meaningfully more tokens and is intended for complex, high-value tasks, not routine edits. The source describes use cases including large migrations, deep research reports, broad fact-checking, backlog triage, post-mortem analysis, security review and design or naming work judged against a rubric.
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 Moves Beyond Solo Agents
The significance of dynamic workflows is that they address limits that can appear when a single agent handles a long or judgment-heavy assignment alone. The Thorsten Meyer AI article identifies three recurring failure modes: agentic laziness, where work stops before all items are handled; self-preferential bias, where an agent grades its own work too favorably; and goal drift, where the original objective weakens over many turns.
By splitting work among separate subagents, Claude can treat one task more like a managed project. One agent may route the work, others may complete narrow pieces, and a separate reviewer may challenge the results. If it works as described, that structure could make AI-assisted work more useful for tasks where parallel execution, independent review and structured synthesis are more valuable than a single continuous conversation.
The tradeoff is cost and control. The source warns that workflows can spawn many agents and consume far more tokens. For readers using AI systems in engineering, research, operations or security review, the practical question is not whether multiple agents sound more powerful, but when the added coordination overhead and token expense are justified.
Agentic AI Engineering: Building AI Agents for Beginners: A Hands-On Guide to No-Code Workflows, LLM Tools, RAG, Automation, and Safe Multi-Agent Systems
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Part Of Claude Code’s Expansion
The new capability fits into a broader pattern of agent orchestration work around Claude Code. According to the source, Anthropic’s recent Claude Code material has described three related building blocks: skills for packaging reusable expertise, loops for deciding how far delegation should continue, and now dynamic workflows for coordinating multiple agents within one task.
The Thorsten Meyer AI article says dynamic workflows compose several recognizable patterns. These include classify-and-act routing, fan-out-and-synthesize work distribution, adversarial verification, generate-and-filter selection, tournaments between competing agents, and loop-until-done structures that continue until a stop condition is met rather than a fixed number of steps.
The source also highlights one security-relevant pattern: quarantine. In that setup, agents that read untrusted public content are kept away from high-privilege actions, while a separate agent performs the action step. The article presents this as a separation-of-duties pattern for autonomous agents, not as a guarantee that every workflow is safe by default.
“The feature is called dynamic workflows, and the plain description is that Claude writes its own harness.”
— Thorsten Meyer AI, summarizing the Claude Code feature
Javascript Cheat Sheet Desk Mat for Software Engineers, Software Development Mouse Mat, Web Developers and Programmers Mouse Pad, Gift Coworker Quick Key, Anti-Slip Keyboard Pad KMH
- Large Desk Space: Fits mouse, keyboard, and accessories
- Smooth Gliding Surface: Ensures precise control and speed
- Portable and Flexible: Rolls up for easy transport
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Reliability Need Proof
Several details remain unclear from the source material. It does not provide independent benchmark results showing how often dynamic workflows outperform a strong single-agent setup, how much token use rises in typical production tasks, or which tasks produce enough quality improvement to justify the added cost.
It is also not clear how users should audit the JavaScript harnesses Claude writes, how workflow failures are surfaced, or what default controls exist around tool access and agent permissions. Anthropic’s framing presents dynamic workflows as useful for complex work, but the available source material does not establish a general performance guarantee across coding, research, security and operations tasks.
Claude at Work for Cloud, DevOps & FinOps: Agentic automation for IT professionals: Claude Code, Cowork, MCP, skills and sub-agents for infrastructure, operations and cloud cost control (2026 Edition)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Users Will Test Boundaries
The next step is adoption and evaluation by teams using Claude Code on large, multi-step work. The source recommends bounding usage with token budgets and pilots before relying on workflows at scale, especially because a poorly constrained workflow could spawn many agents and raise costs quickly.
Readers should watch for more concrete guidance from Anthropic documentation, including examples, limits, model-routing practices and security controls. The practical test will be whether dynamic workflows can produce better verified results on demanding tasks without making cost, review and governance harder to manage.
AI Task Orchestration: Coordinating Complex Agent Workflows. A Comprehensive Guide to Building, Deploying, and Operating Multi-Agent AI Systems
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 a task-specific JavaScript harness to coordinate subagents for complex work.
How are dynamic workflows different from one Claude agent?
A single agent works inside one main context. A dynamic workflow can split a task among multiple subagents, give them focused briefs, wait for their outputs, combine results and assign separate agents to review or challenge the work.
What kinds of tasks are these workflows meant for?
The source points to large migrations, broad research, claim checking, backlog ranking, root-cause analysis, security review and other tasks that are big, parallel, adversarial or judgment-heavy.
Are dynamic workflows better for every request?
No. The source says they use meaningfully more tokens and are aimed at high-value tasks. Routine edits or simple coding requests are unlikely to need a temporary team of agents.
What is still unknown about the feature?
The available source material does not show independent benchmarks, typical cost ranges or detailed default controls for every workflow. It is still unclear how well dynamic workflows perform across different real-world teams and task types.
Source: Thorsten Meyer AI
