TL;DR
Anthropic has described dynamic workflows in Claude Code, a feature that lets Claude generate task-specific JavaScript orchestration to coordinate temporary subagents. The approach is aimed at complex, high-value work where parallel execution, independent review or repeated judging may help, but it can use far more tokens.
Anthropic’s Claude Code team has described dynamic workflows, a feature that lets Claude write a task-specific JavaScript harness to spawn and coordinate subagents for complex work, according to a June 2 Claude blog post by Thariq Shihipar and Sid Bidasaria. The development matters because some Claude Code tasks can now be handled by temporary structured teams rather than one agent working alone.
Under the approach, Claude can generate a JavaScript harness around the model, using special functions to spawn and coordinate subagents while using ordinary JavaScript to handle data. Each subagent can receive a separate context window and a focused task; the source material says subagents may run on different models depending on the job, which remains Anthropic’s implementation description rather than an independent benchmark.
Anthropic’s post, as summarized by Thorsten Meyer AI on July 1, groups the mechanics into six reusable patterns: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament judging and loop-until-done. Those patterns are aimed at work that can be split, checked by a separate agent or repeated until a stop condition is met.
The caveat is cost. The source material says dynamic workflows are built for complex, high-value tasks, not small edits, and warns that workflows can spawn many agents and use far more tokens. That makes token budgets, pilot runs and stop conditions part of the feature’s practical use.
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.
Why Claude’s Agent Teams Matter
Dynamic workflows target failure modes that long agent runs can amplify: partial completion, self-review bias and loss of the original objective after many turns. By assigning separate agents to execution, critique, synthesis and judging, the system attempts a separation-of-duties pattern familiar from human teams.
For software teams, researchers and operations groups, the decision changes from writing a better prompt to deciding whether a task warrants a manager-style workflow. More parallel work can mean more output and more checking, but also higher spend, more latency and more responsibility for workflow design.
JavaScript automation tools for AI workflows
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Skills to Dynamic Workflows
The Thorsten Meyer AI item frames dynamic workflows as the third piece in a loose sequence from the Claude Code team: Skills package organizational knowledge, loops decide how far to delegate over time, and dynamic workflows assemble a temporary team inside a single task.
The examples cited include large migrations and refactors, research reports with citations, claim checking, ticket severity ranking, root-cause postmortems, backlog triage, design or naming by rubric and model routing. The source also highlights a security pattern called quarantine, where agents that read untrusted public content are kept away from high-privilege actions.
“A harness for every task: dynamic workflows in Claude Code”
— Thariq Shihipar and Sid Bidasaria, Anthropic
AI agent orchestration software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs and Guardrails Still Open
The supplied material does not provide independent benchmark results showing lower failure rates, lower review burden or better outcomes than a single-agent run. It is also not yet clear how token overhead, latency and model selection will vary across real teams and task types.
Security details also remain partly about design guidance. The material describes quarantine for untrusted content, with one agent reading risky material and a separate agent taking action, but it does not establish whether Claude Code enforces that split automatically or whether developers must implement the boundary themselves.
complex task management AI tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Developers Test Boundaries Next
Teams evaluating the feature are likely to start with bounded high-value workloads, such as codebase audits, structured research, migration planning or large triage queues. The next practical step is defining token ceilings, stop rules, model-routing choices and human review points before wider use.
Readers should watch for public case studies, pricing impact, admin controls, security examples and updated Claude Code documentation. Those details will show whether dynamic workflows become a routine tool for complex work or stay reserved for large, judgment-heavy tasks.
subagent coordination AI software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are dynamic workflows in Claude Code?
Dynamic workflows are task-specific orchestration programs that Claude can write and run to coordinate multiple subagents. The goal is to split complex work into focused pieces, then merge or review the results.
How is this different from one Claude agent doing the work?
A single agent works inside one main context and may lose track of details during a long task. Dynamic workflows can give subagents separate briefs, separate context windows and roles such as reviewer, judge or synthesizer.
When should teams use dynamic workflows?
The source material points to large, parallel or adversarial tasks, such as migrations, security review, fact-checking, ticket ranking and research reports. It is not presented as the right tool for minor edits or routine prompts.
What are the main risks?
The main tradeoffs are token cost, complexity and the need for clear stopping rules. Security boundaries also matter when subagents read untrusted public content or when any workflow can trigger high-privilege actions.
Has Anthropic proved dynamic workflows work better?
The supplied source describes mechanics and use cases, but it does not include independent performance data. Claims about better reliability should be treated as implementation guidance until more public results are available.
Source: Thorsten Meyer AI
