TL;DR
Anthropic says Claude Code can now create dynamic workflows, writing a JavaScript harness that coordinates temporary subagents for a single complex task. The company presents the approach as useful for parallel, adversarial, or judgment-heavy work, but says it uses more tokens and is not meant for simple requests.
Anthropic says Claude Code can now build dynamic workflows, allowing the model to write a small JavaScript harness that spawns and coordinates subagents for complex tasks, a shift that could change how developers use AI agents for large reviews, migrations, research, and verification work.
The feature, described by Anthropic in a Claude blog post titled A harness for every task: dynamic workflows in Claude Code, lets Claude create orchestration scaffolding around a task instead of relying on a single agent to plan, execute, and check its own work. According to the source material, the workflow can assign separate roles, wait for results, merge structured outputs, and use independent agents to review or challenge the work.
Thorsten Meyer AI framed the change as Claude drawing an org chart for one job: a dispatcher, specialists, reviewers, and judges that exist only for the task at hand. The confirmed mechanism is narrower and more technical: Claude writes a JavaScript program with functions for spawning and coordinating subagents, while using ordinary JavaScript to handle data and control flow.
Anthropic’s caveat is central to the announcement. Dynamic workflows are described as suitable for complex, high-value tasks, not small edits or routine prompts. The source material says the approach can use meaningfully more tokens, and workflows may spawn many agents if they are not bounded by budgets, pilots, or stopping 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.
When Claude Needs A Team
The development matters because it targets known weaknesses in long agent runs: partial completion, self-review bias, and goal drift over extended tasks. Instead of asking one model instance to handle everything, the workflow can split work across isolated subagents, then have another agent verify the results.
That could make Claude Code more useful for tasks where parallel work and independent checking are more valuable than a single conversational thread. The cited use cases include large refactors, deep research reports, claim-by-claim fact checks, ticket ranking, post-mortem analysis, backlog triage, design judging, model routing, and security review patterns.
The tradeoff is cost and management. Anthropic’s framing makes clear that dynamic workflows do not remove the need for human judgment. They shift the user’s role from prompting one worker to commissioning a temporary agent team, which can produce more coverage but also consume more compute, tokens, and review time.
JavaScript development tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Anthropic’s Agent Workflow Arc
The announcement sits inside a broader Claude Code push described by Thorsten Meyer AI as a three-part arc. Skills package organizational knowledge, loops decide how far work is delegated over time, and dynamic workflows coordinate several subagents inside one task.
The pattern borrows from common engineering and management practices: divide a large job, isolate responsibilities, merge outputs, and send results to an independent reviewer. In Anthropic’s description, subagents can have their own context windows, focused briefs, and in some cases different model choices, allowing the system to avoid overloading one context with every detail of a large assignment.
The source material lists six workflow patterns: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournaments, and loop-until-done. These are presented as composable moves Claude can use when the task benefits from routing, parallel work, judging, or repeated attempts under a stop condition.
“Claude writes its own harness”
— Thorsten Meyer AI
AI workflow automation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Limits Still Open
Several details remain unresolved from the supplied material. It is not yet clear how often Claude Code will choose dynamic workflows in ordinary use, how users will set practical limits, or how predictable token costs will be when workflows spawn many agents.
Performance claims also need careful reading. The source material describes likely benefits for complex tasks, but it does not provide independent benchmark results showing how often dynamic workflows outperform single-agent runs across real projects. It also does not specify failure rates, default safety limits, or the exact model-routing behavior across subagents.
Security handling is another area to watch. The source material highlights a quarantine pattern: agents that read untrusted public content should be barred from high-privilege actions, while a separate agent acts on verified information. That is a defensive design pattern, but the supplied material does not confirm how much of it is enforced by product defaults versus left to workflow design.
subagent management tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Developers Test The Boundaries
The next step is adoption by Claude Code users on work that is large enough to justify the added cost. Teams are likely to test dynamic workflows first on bounded pilots, such as a refactor review, a research memo, a ticket-ranking job, or a claim-checking pass where the output can be compared against a rubric.
Anthropic’s documentation at code.claude.com/docs is the place users are directed for implementation details. Readers should watch for clearer guidance on token budgets, stop conditions, model selection, safety defaults, and examples showing when a workflow beats a single-agent prompt.
AI orchestration platforms
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 for Claude Code, a capability where Claude writes and runs a task-specific JavaScript harness to coordinate temporary subagents.
Is this meant for everyday coding tasks?
No. The source material says the feature is aimed at complex, high-value work. Anthropic’s caveat is that it can use more tokens, so it is not suited to minor edits or simple prompts.
How is a dynamic workflow different from one agent?
A single agent works inside one main context. A dynamic workflow can split a task among separate subagents, give them focused goals, wait for their outputs, and use independent review to check results.
What kinds of tasks could benefit?
The cited examples include large migrations, refactors, deep research, fact-checking, ticket ranking, post-mortems, backlog triage, design judging, and security review patterns.
What remains uncertain?
Open questions include real-world cost, default safety controls, how often workflows outperform single-agent runs, and how users should set budgets and stopping rules.
Source: Thorsten Meyer AI
