TL;DR
Anthropic says Claude Code can now create dynamic workflows, writing a task-specific JavaScript harness that coordinates temporary subagents. The approach is aimed at large, high-value tasks, but the company says it uses meaningfully more tokens and is not suited to simple work.
Anthropic says Claude Code can now build dynamic workflows for complex tasks, writing a temporary JavaScript harness that coordinates multiple subagents instead of relying on one agent to handle the full job alone.
The capability, described by Anthropic in a June 2 post titled “A harness for every task: dynamic workflows in Claude Code”, lets Claude create the orchestration scaffolding around a task. According to the source material, that harness can spawn subagents, assign them separate context windows, wait for their outputs, and merge results into a final answer.
Anthropic’s framing is narrower than a general productivity claim. The company says the pattern is built for complex, high-value work and uses meaningfully more tokens. In plain terms, it is meant for jobs where separate workers, independent review, or parallel handling can reduce errors, not for small edits or simple requests.
The workflows can combine several patterns, including classify-and-act routing, fan-out-and-synthesize, adversarial checking, generate-and-filter selection, tournament-style judging, and loop-until-done execution. Those patterns remain claims about intended behavior from Anthropic and Thorsten Meyer AI’s summary; the source material does not provide independent benchmark results for each use case.
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.
Complex Tasks Get More Structure
The development matters because many agent failures come from asking one model context to plan, execute, review, and remember a large assignment. The source material points to three recurring problems: stopping early, favoring its own output, and losing the original objective over long sessions.
Dynamic workflows are meant to address those risks by separating roles. A subagent can handle a narrow brief, another can review the output, and an orchestrator can combine the results. For teams using AI on large code changes, research reports, backlog triage, fact checks, or security reviews, that separation may make agent work easier to audit.
The trade-off is cost and control. Anthropic’s own caveat, as cited in the source material, is that the method consumes more tokens. That makes task selection central: a workflow can make sense for parallel or judgment-heavy work, while a single-agent request remains better for routine jobs.
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.
Claude Code’s Third Workflow Layer
Thorsten Meyer AI frames the announcement as the third part of a loose arc from the Claude Code team. In that framing, skills package organizational knowledge, loops decide how long to keep delegating work, and dynamic workflows decide how to assemble a team inside one task.
The source material says the workflow is mechanically a small JavaScript program written and run by Claude. That program can use special functions for subagent coordination while using ordinary JavaScript to handle data, merge outputs, and manage the task flow.
A security point in the source material is role separation. Agents that read untrusted public content should be kept away from high-privilege actions, while a different agent performs the action. That pattern is presented as a way to reduce the risk that external content influences an agent with broader permissions.
“Claude writes its own harness”
— Thorsten Meyer AI
Javascript & Data Structures Study Flashcards – Master Core Concepts, Algorithms, and Interview-Ready Developer Skills
Comprehensive Coverage: Dive deep into JavaScript with thorough explanations of key topics and practical, real-world examples that make…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Limits Still Need Proof
It is not yet clear how often dynamic workflows outperform a simpler single-agent process in live production settings. The source material lists potential use cases, but it does not include independent measurements of accuracy gains, token cost per task, latency, or failure rates.
It is also unclear how developers should set practical limits. The article warns that workflows can spawn many agents and burn far more tokens, but it does not give a universal threshold for when to use one agent, a small workflow, or a large multi-agent process.
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.
Developers Will Test Boundaries
The next step is likely hands-on evaluation by Claude Code users. Teams will need to test dynamic workflows on bounded tasks, compare results against simpler agent runs, and decide where the added cost is justified.
Anthropic’s documentation at code.claude.com/docs is the stated place to watch for implementation guidance. For now, the clearest recommendation from the source material is to start with pilot workflows, set token budgets, and reserve the pattern for work that is large, parallel, adversarial, or review-heavy.
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 where Claude writes a task-specific JavaScript harness to spawn and coordinate subagents during one complex job.
Is this meant for everyday coding tasks?
No. The source material says the approach uses meaningfully more tokens and is aimed at complex, high-value tasks, not simple fixes or small edits.
What kinds of work could use dynamic workflows?
The cited examples include large migrations, deep research reports, claim checking, backlog ranking, root-cause analysis, security review, naming work by rubric, and model routing.
What remains unknown about the feature?
The main open questions are real-world cost, reliability, latency, and how often a multi-agent workflow beats a well-scoped single-agent run on the same task.
Source: Thorsten Meyer AI
