TL;DR
Anthropic says Claude Code can now generate dynamic workflows that coordinate specialized subagents within a single task. The feature is aimed at complex work where one agent may miss items, favor its own output, or lose the original goal, but it also uses more tokens and still leaves open questions about cost and controls.
Anthropic says Claude Code can now create dynamic workflows that spawn specialized subagents inside a single task, a change meant to help developers and teams handle complex AI work that can overwhelm one agent.
The feature, described by Anthropic in a June 2, 2026 Claude blog post titled A harness for every task: dynamic workflows in Claude Code, lets Claude write a small JavaScript harness for the task at hand. That harness can spawn subagents, coordinate their work, wait for returns at a barrier, and merge structured outputs into a final answer, according to the report.
The mechanics are meant to address three failure modes identified in the source material: agentic laziness, self-preferential bias, and goal drift. Instead of leaving one model to plan, execute, and grade its own work, Claude can give isolated agents narrower briefs and ask separate reviewers to test the output.
The system can compose several patterns, including classify-and-act routing, fan-out-and-synthesize, adversarial verification, and tournament-style judging. Anthropic’s caveat, repeated in the report, is that these workflows use more tokens and are aimed at complex, high-value tasks rather than 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.
Agent Teams Target Hard Tasks
For software teams, the practical value is not that Claude has more personas. It is that parallel work, independent review, and structured synthesis can be built into one run when the job is too broad for a single context window.
The reported use cases include large migrations, security reviews, claim-by-claim fact checks, and ranking big ticket backlogs. Those are jobs where partial completion, unchecked self-review, or lost instructions can turn a fast AI task into a costly cleanup job.
The tradeoff is cost and supervision. If a workflow can spawn many agents, teams need token budgets, pilot runs, and clear stop conditions before using it on production work or time-sensitive research.
Agentic AI Engineering: Building AI Agents for Beginners: A Hands-On Guide to No-Code Workflows, LLM Tools, RAG, Automation, and Safe Multi-Agent Systems
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Claude Code’s Orchestration Layer
The Thorsten Meyer AI report places dynamic workflows after two other Claude Code ideas: Skills, which package organizational knowledge, and loops, which decide how long to keep delegating over time. Dynamic workflows cover a different axis: one task with a temporary set of agents.
That puts the feature in the same broader move toward agent orchestration, where a main model manages workers, reviewers, or judges. The new element described here is on-the-fly harness generation, not merely a fixed checklist or static prompt template.
“A harness for every task”
— Thariq Shihipar and Sid Bidasaria, Anthropic Claude blog
The Human-Agent Orchestrator: Leading and Scaling AI-Driven Organizations
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Controls Remain Open
It is not yet clear how often dynamic workflows will produce better results than a strong single-agent run in ordinary teams. The source material describes the feature as recent and still evolving, and it does not provide public benchmark results for the full range of tasks named in the report.
Several operational details also remain developing: rollout scope, default limits, model routing choices, and how teams should audit outputs when many subagents contribute. The report also flags a security pattern called quarantine, where agents that read untrusted public content are kept away from high-privilege actions.
Building Agentic AI Systems with Claude Code: A Practical Guide to MCP, Sub-Agents, Hooks, Computer Use, and Production-Ready AI Workflows in 2026
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Developers Test The Harnesses
The next step is practical evaluation inside Claude Code projects. Teams that adopt the pattern are likely to start with bounded pilots, token caps, and review checkpoints on tasks where parallelism or adversarial review is useful enough to justify the added cost.
Readers should watch Anthropic’s Claude Code docs and future Claude blog posts for clearer guidance on availability, security controls, and recommended workflow limits. For now, the clearest rule from the source material is simple: use dynamic workflows for work that truly needs a team.
AI task coordination platform
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Anthropic change in Claude Code?
Anthropic says Claude Code can create dynamic workflows, meaning it can generate a task-specific harness that coordinates specialized subagents during one job.
Does Claude literally hire agents?
No. The report uses the team metaphor to describe software orchestration: separate subagents get focused briefs, isolated context, and structured outputs that are merged later.
Is this meant for everyday edits?
No. The stated caveat is higher token use. The feature is aimed at complex tasks such as migrations, security checks, research reports, or large-scale triage.
What problems is this supposed to reduce?
The report points to early stopping, self-review bias, and goal drift. Dynamic workflows try to reduce those problems by splitting work and adding independent checks.
What remains uncertain for users?
Open questions include cost in real deployments, default guardrails, and how much accuracy improves across different task types outside the examples described by Anthropic and Thorsten Meyer AI.
Source: Thorsten Meyer AI
