TL;DR
Anthropic’s Claude Code team has described dynamic workflows, a feature that lets Claude write a task-specific JavaScript harness to spawn and coordinate temporary subagents. The company frames it as a way to split complex, parallel, or review-heavy work, while warning that it can use far more tokens.
Anthropic says Claude Code can now use dynamic workflows to write and run task-specific JavaScript harnesses that coordinate temporary subagents, a capability aimed at complex work where one agent may miss items, favor its own output, or lose the original goal.
The feature, as described by Anthropic and summarized by Thorsten Meyer AI, lets Claude create the orchestration scaffolding around a task. That harness can spawn subagents with separate context windows, focused instructions, and, in some cases, different models or worktrees.
The source material says Claude can combine several workflow patterns, including classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament judging, and loop-until-done execution. In practice, that means Claude can divide a large task, wait for separate agents to finish, merge structured outputs, and send results to an independent checker.
Anthropic’s stated caveat is cost and scope. The approach uses meaningfully more tokens and is intended for complex, high-value tasks, not routine edits. The Thorsten Meyer AI dispatch warns that workflows can become expensive if teams do not set budgets, run small pilots, and define stop conditions.
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.
Agent Teams for Hard Tasks
The development matters because many AI-agent failures come from putting too much work inside one context window. The source material identifies three recurring problems: agentic laziness, self-preferential bias, and goal drift during long tasks.
Dynamic workflows are meant to address those problems by separating roles. A planner can assign work, specialists can handle narrow parts, and an independent reviewer can challenge the result. According to the source material, that could help with large refactors, deep research reports, fact-checking, ticket ranking, root-cause reviews, backlog triage, design selection, and model routing.
The trade-off is that users move from prompting one worker to commissioning a temporary team. That can produce more coverage and review, but it also adds token cost, orchestration complexity, and a need for clear manager-style judgment.
AI workflow automation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Claude Code’s Third Agent Pattern
The Thorsten Meyer AI dispatch frames dynamic workflows as the third part of a loose arc from Anthropic’s Claude Code team. In that framing, Skills package an organization’s knowledge, loops decide how far to delegate over time, and dynamic workflows decide how to split work inside a single task.
The Anthropic source cited by the dispatch is A harness for every task: dynamic workflows in Claude Code, credited to Thariq Shihipar and Sid Bidasaria and dated June 2, 2026. The July 1, 2026 Thorsten Meyer AI piece translates that technical idea into an org-chart analogy: Claude assembles a short-lived team, gives each agent a desk and brief, and disbands the group when the work is done.
“Dynamic workflows are the third axis.”
— Thorsten Meyer AI dispatch
JavaScript harness development kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs and Results Still Open
The source material does not provide independent benchmark data showing how much dynamic workflows improve task completion, review quality, or cost efficiency across real projects. It is also not clear from the provided material how broadly the feature is available, what limits Anthropic applies, or how users should compare one-agent and multi-agent runs.
Security behavior is another developing area. The dispatch highlights a quarantine pattern: agents that read untrusted public content should be blocked from high-privilege actions, while a separate agent handles acting. The provided material does not give a full enforcement model for that separation.
multi-agent AI system
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Docs and Adoption to Watch
The next milestone is practical adoption through Claude Code documentation and controlled experiments by teams working on large agent tasks. Users evaluating the feature will need to compare accuracy, cost, latency, and review quality against simpler single-agent workflows.
For now, the clearest use cases are bounded tasks that are large, parallel, adversarial, or judgment-heavy. Routine coding edits, typo fixes, and small one-step requests remain poor fits for a multi-agent workflow.
AI task orchestration software
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 harnesses that Claude can write and run to coordinate temporary subagents. They let Claude split a complex job into focused pieces and merge the results.
Is this meant for everyday coding tasks?
No. The source material says the approach is built for complex, high-value work and uses more tokens. Small edits and routine fixes are better handled by a single agent.
How are subagents different from one Claude session?
Each subagent can receive a focused brief and work in a separate context window. That separation is meant to reduce overload, self-review bias, and loss of the original task goal.
What risks come with this approach?
The main risks are higher cost, added orchestration overhead, and poorly bounded runs. The source material says teams should set budgets, run small pilots, and define stop conditions.
Where could teams use it first?
The source material points to big migrations, refactors, research reports, fact-checking, ticket triage, post-mortems, and security-style review tasks where parallel work and independent checking matter.
Source: Thorsten Meyer AI
