TL;DR
Anthropic has described dynamic workflows in Claude Code, a system that lets Claude write task-specific JavaScript orchestration code and coordinate multiple subagents. The company says the approach is meant for complex, high-value tasks, while cost, reliability and limits remain developing questions.
Anthropic has described a new Claude Code capability called dynamic workflows, in which Claude can write orchestration code during a task and coordinate multiple temporary subagents for complex work, a development aimed at reducing failures that can arise when one agent handles a large job alone.
The feature was described by Thariq Shihipar and Sid Bidasaria in Anthropic’s Claude blog post, A harness for every task: dynamic workflows in Claude Code, published June 2, 2026. According to the account, Claude can create a small JavaScript harness that routes work, launches subagents, waits for results and merges structured outputs.
Thorsten Meyer AI, writing on July 1, 2026, framed the change as the third part of a loose arc from the Claude Code team: skills package organizational knowledge, loops decide how delegation proceeds over time, and dynamic workflows decide how several agents cooperate within one task. That framing is the site’s interpretation; the mechanics and examples are attributed to Anthropic.
Anthropic’s own caveat is material: this approach uses meaningfully more tokens and is intended for complex, high-value tasks, not routine edits. The company’s described patterns include routing by task type, parallel fan-out and synthesis, adversarial review, generate-and-filter work, tournaments and looped delegation 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.
Claude Moves From Worker To Coordinator
The development matters because it changes the model of agentic work from a single assistant completing a long task to a system that can assign separate briefs to separate subagents. In theory, that can reduce problems such as partial completion, weak self-review and loss of the original objective during long sessions.
For readers using AI tools at work, the practical effect is clearer in tasks such as large migrations, deep research, security reviews, backlog triage and root-cause analysis. These are jobs where parallel work, independent checking and structured synthesis can matter more than one agent producing a fast answer.
The tradeoff is cost and control. Anthropic and Thorsten Meyer AI both stress that dynamic workflows can burn far more tokens, and that users need token budgets, pilot runs and clear stopping rules before sending a model to spawn many agents.
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.
Why Single Agents Can Fail
The feature is designed around well-known agent failure modes. The source material identifies agentic laziness, where an agent declares a task complete after partial work; self-preferential bias, where it favors its own output during review; and goal drift, where the original request weakens across long sessions.
Dynamic workflows address those problems by separating roles. One agent may orchestrate, others may perform focused work, and another may run adversarial verification. Anthropic’s described approach gives subagents clean context windows and focused goals, rather than forcing all steps into one overloaded thread.
The source material also flags a security pattern: quarantine. Agents that read untrusted public content should be barred from privileged actions, while a separate agent handles action-taking. That separation is presented as a way to reduce risk when autonomous systems touch outside information.
“A harness for every task”
— Anthropic blog post by Thariq Shihipar and Sid Bidasaria
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.
Cost And Reliability Questions Remain
Several details remain unclear from the source material. It is not yet clear how broadly dynamic workflows are available across Claude Code users, what limits Anthropic applies to subagent spawning, or how often the approach improves results enough to justify the added token cost.
It is also unclear how users should measure reliability across many workflow types. Anthropic describes patterns and use cases, but the available material does not provide independent benchmark results showing when multi-agent orchestration consistently beats a well-prompted single-agent workflow.
ENTERPRISE MULTI-AGENT AI SYSTEMS: Scaling Collaborative Agent Swarms, Reasoning Frameworks, and Cross-Functional AI Teams (The Enterprise AI Architect’s Handbook)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Teams Will Test Workflow Boundaries
The next step is practical evaluation. Developers and technical teams are likely to test dynamic workflows on bounded tasks such as refactors, claim-checking, ticket ranking and post-mortem analysis, where outputs can be reviewed against known criteria.
Anthropic’s documentation at code.claude.com/docs is the place to watch for implementation details, limits and usage guidance. For now, the confirmed message is narrow: dynamic workflows are aimed at large, parallel or judgment-heavy tasks, while routine work still calls for a simpler agent setup.
Optimising Project Workflows with AI – Practical Guide: Powered by Agentic AI + Make.com
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 create task-specific orchestration code and coordinate multiple subagents during a complex task.
Is this meant for everyday Claude tasks?
No. The source material says the approach uses more tokens and is built for complex, high-value work, not simple edits or routine requests.
What kinds of work could use dynamic workflows?
Examples cited include large migrations, deep research, fact-checking, security review, backlog triage, design evaluation and root-cause analysis.
What remains unknown?
Availability, practical limits, cost controls and independent evidence of reliability gains remain unclear or still developing based on the supplied material.
Source: Thorsten Meyer AI
