TL;DR
Anthropic says Claude Code can write a JavaScript harness that spawns and coordinates subagents for one complex task. The feature, called dynamic workflows, is aimed at high-value work because it can use far more tokens than a single-agent run.
Anthropic’s Claude Code team has introduced dynamic workflows, a feature that lets Claude write a JavaScript harness to spawn and coordinate task-specific subagents during one complex job, according to Anthropic material cited by Thorsten Meyer AI. The development matters because it moves Claude Code beyond a single-agent setup for large refactors, research checks and security reviews, while raising cost and oversight questions.
The confirmed development is narrower than the shorthand suggests: Claude is not adding permanent coworkers. Anthropic describes dynamic workflows as task-specific orchestration code that Claude writes and runs for the job in front of it, using special functions to spawn subagents, pass focused briefs, wait at a barrier and merge outputs. Thorsten Meyer AI summarized the pattern as Claude drawing an org chart for one job, then disbanding that temporary team when the work is done.
According to the supplied source material, each subagent can receive a clean context window, a focused goal and, in coding work, an isolated workspace. The approach can also route work to different model choices, allowing a task to use cheaper or stronger models where the harness calls for them. Anthropic’s named workflow moves include fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament judging and loop-until-done execution.
The key caveat is cost. The source material says Anthropic presents dynamic workflows for complex, high-value tasks, not small edits, because the method can use meaningfully more tokens. Claims that the approach reduces early stopping, self-grading bias and goal drift are attributed to Anthropic’s workflow framing and the Thorsten Meyer AI write-up; the supplied material does not provide independent benchmark results.
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 Costlier Team Model
For software teams, the new model changes how AI-assisted work may be scoped. A single Claude instance can be asked to act less like one worker with a long checklist and more like a temporary project structure with specialists, reviewers and judges. That could matter for large migrations, security reviews, research reports and backlog triage, where parallel work and independent checking are part of the job.
The trade-off is that dynamic workflows bring manager-like decisions into everyday AI use: when to split work, how many agents to run, what stop condition to set and how to cap token spending. Thorsten Meyer AI’s reading is that users should reach for the feature when a task is big, parallel, adversarial or judgment-heavy. For routine fixes, the same source warns that a multi-agent workflow can be wasteful.
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 Axis
Thorsten Meyer AI frames dynamic workflows as the third piece in a sequence from Anthropic’s Claude Code team. In that framing, skills package an organization’s know-how, loops decide how far to delegate over time and dynamic workflows coordinate several agents within one task.
The development also reflects a wider push in agent tooling toward orchestration rather than one long prompt. The source material lists use cases beyond code, including deep research reports, claim-by-claim fact checks, ranking 1,000 tickets and root-cause reviews after business metrics move unexpectedly.
“A harness for every task: dynamic workflows in Claude Code”
— Thariq Shihipar and Sid Bidasaria, Anthropic
Hierarchical Sub-Agent Orchestration in CrewAI and LangGraph: Optimizing task delegation and state-persistence to reduce token overhead and prevent agentic loops
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Access Remain Open
It is not yet clear from the supplied material how broadly dynamic workflows are available across Claude Code users, how token budgets will be surfaced in every environment or what defaults limit runaway spawning. The material also leaves open how much quality gain teams can expect on real workloads, because it cites patterns and use cases rather than public comparative results.
Security details are also developing. Thorsten Meyer AI highlights a quarantine pattern: agents that read untrusted public content should be barred from high-privilege actions, with a separate agent handling the action step. The available source material does not establish how fully that separation is enforced by Claude Code itself versus by a user’s harness design.
40 Python Programming Projects for AI Agents and Automations: Build Agentic Systems, Bots, and Automated Tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Docs And Pilot Runs
Teams evaluating the feature should watch Anthropic’s Claude Code documentation and run bounded pilots before routing costly work through a dynamic workflow. Practical next steps include setting token caps, starting with a small representative task, comparing results against a single-agent run and logging when subagents add value.
The next milestone for readers is evidence from real deployments: whether multi-agent orchestration reliably improves accuracy, coverage and review quality enough to justify higher cost. Until more results are public, dynamic workflows are best treated as a specialized tool for high-stakes, complex jobs, not a default for every Claude request.
No-Code AI Marketing Workflows with Zapier and Make: Automate Leads, Content, Follow-Up, and Campaigns Without Coding Using AI-Powered Systems (AI Toolkit For Online Marketers Book 17)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Anthropic announce for Claude Code?
Anthropic described dynamic workflows, a Claude Code capability in which Claude writes a JavaScript orchestration harness that can spawn, coordinate and merge work from subagents during one task.
Are dynamic workflows permanent agent teams?
No. Based on the supplied material, the subagents are temporary workers for one job. Claude creates the task-specific structure, waits for the agents’ outputs and folds the results into a final answer.
Why not use one Claude agent?
The claim from Anthropic’s workflow framing is that one long-running agent can stop early, favor its own output or lose the original goal. Splitting work into isolated briefs and adding independent review is meant to reduce those failure modes on larger tasks.
Does this make Claude Code cheaper?
No. The caveat in the source material is the opposite: dynamic workflows can use far more tokens. They are aimed at complex, high-value tasks where extra checking or parallel work may be worth the added cost.
What should teams verify before using it?
Teams should verify access, pricing, token controls and safety boundaries in Anthropic’s official documentation before using the feature for production work. They should also test whether subagent results beat a simpler single-agent run on their own tasks.
Source: Thorsten Meyer AI
