TL;DR
Anthropic has described dynamic workflows in Claude Code, a capability in which Claude writes a JavaScript harness to spawn and coordinate task-specific subagents. The company frames the approach as useful for complex, high-value work, while Thorsten Meyer AI highlights the token cost and need for management judgment.
Anthropic’s Claude Code can now use dynamic workflows that let Claude write task-specific JavaScript orchestration code to coordinate temporary subagents, a development aimed at complex tasks where one agent may miss work, favor its own output or lose the original goal.
In a June 2, 2026 Claude blog post titled A harness for every task: dynamic workflows in Claude Code, Thariq Shihipar and Sid Bidasaria described a workflow system that Claude can generate for the job at hand. The harness can fan work out to separate agents, wait for their outputs and merge the results into one answer.
The Thorsten Meyer AI dispatch published July 1, 2026 describes the approach as Claude drawing an org chart for a single job, with roles such as a dispatcher, specialists, an independent reviewer or a panel of judges. That framing is the author’s; the mechanics, patterns and use cases are attributed to Anthropic.
The documented patterns include classify-and-act routing, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournaments and loop-until-done workflows. Anthropic’s caveat, as relayed by the dispatch, is that this uses meaningfully more tokens and is meant for complex, high-value tasks, not small edits.
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.
Multi-Agent Work For Hard Tasks
The change matters because single-agent work can degrade on long or adversarial jobs: the source material cites early stopping, self-grading bias and goal drift as common failure modes. A dynamic workflow can assign narrower briefs, keep each subagent’s context cleaner and use independent review before a final synthesis.
For readers using AI tools at work, the appeal is practical rather than theatrical: possible use cases include large refactors, security reviews, fact-checking, research reports, ticket ranking and root-cause post-mortems. The trade-off is cost and control, since a workflow can spawn many agents and consume far more tokens.
Javascript of Automation: Crafting Scripts for macOS (Aquitaine Programming)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Claude Code’s Workflow Stack
The feature sits after two related Claude Code ideas described in the source material: skills, which package organization knowledge, and loops, which decide how far delegation continues over time. Dynamic workflows add a third layer inside one task: custom orchestration created for the current job.
Anthropic’s examples are not limited to coding. The dispatch lists deep research into a cited report, ranking 1,000 tickets by severity, triaging a backlog, design and naming by rubric, and model routing. It also highlights a security pattern: quarantine, where agents that read untrusted public content cannot take high-privilege actions.
“A harness for every task: dynamic workflows in Claude Code”
— Thariq Shihipar and Sid Bidasaria, Anthropic
ENGINEERING PLANNING & ORCHESTRATION : The AI-Native Quartet: RAG · Knowledge Graphs · Context & Memory · Planning (The AI-Native Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Boundaries Remain Open
Several details are still developing. The source material does not give benchmarks showing when dynamic workflows outperform a single agent, nor does it quantify token budgets, latency or error rates across task types. It is also not clear how often teams will accept the added cost for routine work.
The security model also needs careful implementation. The dispatch identifies separation of duties and quarantine as a pattern for untrusted content, but public details in the supplied source do not establish how every deployment will enforce those boundaries.
Designing Multi-Agent Systems: Principles, Patterns, and Implementation for AI Agents
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Teams Test Workflow Guardrails Next
The next phase is adoption and evaluation. Developers are likely to test pilot workflows on bounded tasks, set token caps, compare outputs against single-agent baselines and decide which jobs need subagents, reviewers or judges.
Readers following the feature should watch Claude Code documentation and Anthropic engineering posts for clearer guidance on availability, model routing and safeguards. For now, the confirmed takeaway is narrow: dynamic workflows give Claude Code a way to assemble a task-specific team, with cost and governance questions still open.
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 in which Claude writes a small JavaScript harness to coordinate subagents for a specific task.
Is this the same as using one Claude agent with a long prompt?
No. A long prompt keeps work inside one main context, while dynamic workflows split work among separate agents with focused briefs and then merge results through synthesis.
When should teams use dynamic workflows?
The supplied source points to large refactors, research reports, security reviews, backlog triage and judgment-heavy tasks. It also says the approach is a poor fit for small edits because of token cost.
What are the main risks?
The main risks are higher token use, added orchestration complexity and weak boundaries if agents handling untrusted content can also take privileged actions. The dispatch highlights quarantine as a mitigation pattern.
Is the feature fully proven?
The feature is described as recent and still evolving. The source does not provide comparative benchmarks, so the strongest claims about productivity and quality remain claims to test in real use.
Source: Thorsten Meyer AI
