TL;DR
Anthropic has detailed dynamic workflows for Claude Code, a feature that lets Claude write a task-specific JavaScript harness and coordinate temporary subagents. The company frames it as useful for complex, high-value work, while the main open issues are cost, reliability and where teams should draw the line.
Anthropic’s Claude Code team has described dynamic workflows, a system in which Claude writes a JavaScript orchestration harness and coordinates temporary subagents for complex tasks, a change that could widen what agentic coding tools can handle while using more tokens.
According to the source material, a dynamic workflow is a small program Claude writes for the job in front of it. That program can spawn subagents, assign them focused briefs, keep their work isolated and then merge their structured outputs into one final result.
The mechanics described by Anthropic include agents with separate context windows, possible use of different model tiers, parallel execution and an independent review step. The source says this design is meant to reduce common single-agent failure modes, including partial completion, self-grading bias and drift from the original goal.
Anthropic’s caveat is direct: the approach uses meaningfully more tokens and is intended for complex, high-value tasks, not small edits. The workflows can combine patterns such as fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament-style judging and loop-until-done execution.
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 Manager Layer for Claude
The development matters because it moves Claude Code from a single-agent model toward a managed team pattern. For developers, researchers and operations teams, that could make large migrations, broad fact-checking, backlog triage and security reviews easier to split across specialized agents.
The claimed benefit is not that every answer becomes better. The more specific claim is that independent subagents can reduce overload, check one another’s work and keep separate parts of a task from contaminating the same context window.
The tradeoff is cost and control. If a workflow can spawn many agents, teams need token budgets, stop conditions and review steps before using it on production work or high-stakes decisions.
AI agent orchestration software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Skills to Workflows
The Thorsten Meyer AI dispatch frames dynamic workflows as the third part of a Claude Code arc: skills package organizational knowledge, loops decide how far to delegate over time and dynamic workflows coordinate several agents inside one task.
The dispatch cites Anthropic’s June 2, 2026 Claude blog post, A harness for every task: dynamic workflows in Claude Code, by Thariq Shihipar and Sid Bidasaria. The source attributes the underlying mechanics, patterns and use cases to Anthropic, while the “org chart” framing comes from the dispatch author.
“A harness for every task: dynamic workflows in Claude Code”
— Thariq Shihipar and Sid Bidasaria, Anthropic
JavaScript AI task harness tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs and Reliability Remain Open
It is not yet clear how often Claude will choose the right number of agents, the right model for each role or the right stopping point without human guardrails. Public source material does not give benchmark results showing how dynamic workflows perform across common enterprise tasks.
Availability details are also limited in the provided material. The dispatch points readers to code.claude.com/docs, but the rollout status, pricing impact and enterprise controls may vary by Claude Code environment.
AI subagent management system
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Pilots, Budgets and Guardrails
The next step for teams is likely limited testing on tasks that are large, parallel, adversarial or judgment-heavy. The dispatch recommends starting with bounded pilots, explicit token limits and checks that compare workflow output against single-agent results.
Security practices will also matter. The source highlights a quarantine pattern: agents that read untrusted public content should not receive high-privilege tools, while a separate agent handles privileged actions.
complex AI workflow automation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are Claude Code dynamic workflows?
Dynamic workflows are task-specific orchestration programs that Claude writes to coordinate temporary subagents, assign focused work and synthesize the results.
Is this meant for everyday coding edits?
No. The source material says the feature uses more tokens and is aimed at complex, high-value work, not small fixes such as a typo.
Why use several agents instead of one?
The stated reason is to reduce single-agent failure modes, including early stopping, self-review bias and goal drift during long tasks.
What kinds of tasks fit this model?
The source lists big migrations, deep research, claim-checking, ticket ranking, post-mortems, backlog triage and security review patterns as likely fits.
What should teams watch before using it?
Teams should watch token spend, agent count, stop conditions, tool permissions and whether an independent reviewer catches workflow errors before results are trusted.
Source: Thorsten Meyer AI
