TL;DR
Anthropic’s Claude Code team has described dynamic workflows, a feature that lets Claude write a JavaScript harness and coordinate temporary subagents during a single complex task. The company says the approach can reduce single-agent failure modes, but it uses more tokens and is aimed at high-value work rather than routine edits.
Anthropic’s Claude Code can now use dynamic workflows that let Claude write task-specific orchestration code and coordinate a temporary team of subagents, according to a June 2 company blog post by Thariq Shihipar and Sid Bidasaria. The development matters because it moves some complex AI work from a single long-running agent toward a managed, multi-agent process with separate roles, checks and synthesis.
The feature, described by Anthropic as “dynamic workflows”, lets Claude create a small JavaScript harness for the task at hand. That harness can spawn subagents, assign focused briefs, wait for their work, compare outputs and merge results. Thorsten Meyer AI framed the change as Claude drawing an org chart for one job, then disbanding that team when the work is complete.
Anthropic’s stated patterns include classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournaments and loop-until-done workflows. In practical terms, one subagent might research one part of a problem, another might test a claim, and a separate reviewer might challenge the result before the orchestrator produces a final answer.
The company’s caveat is central: this approach uses meaningfully more tokens and is meant for complex, high-value tasks. The source material says it is not the right method for small edits, such as fixing a typo, because the cost and coordination overhead can outweigh the benefit.
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.
Single-Agent Limits Get Exposed
The announcement matters because it addresses a known weakness in long agentic workflows: one agent can lose track of the original goal, stop after partial progress or judge its own work too favorably. Thorsten Meyer AI describes those failure modes as agentic laziness, self-preferential bias and goal drift.
Dynamic workflows try to reduce those risks by separating responsibilities. A subagent working in a clean context window can focus on a narrow assignment, while another agent can review the output without having produced it. That separation is especially relevant for work such as large code migrations, security reviews, deep research, ticket triage and root-cause analysis.
The tradeoff is cost and control. A workflow that spawns many agents may produce better coverage, but it can also consume far more compute and tokens. Readers using Claude Code should treat the feature less like a default mode and more like a tool for tasks that are big, parallel, adversarial or judgment-heavy.
JavaScript harness for AI workflows
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Part Of Claude Code’s Arc
The Thorsten Meyer AI piece places dynamic workflows after two related Claude Code ideas: skills, which package organizational knowledge, and loops, which decide how work should be delegated over time. Dynamic workflows add a third dimension: delegation inside a single task.
Under the hood, the key shift is not only that Claude calls other agents. It is that Claude can write the orchestration scaffold around the model for the job in front of it. That makes the workflow more flexible than a fixed script, while still letting developers apply boundaries such as budgets, stop conditions and review steps.
The source material also highlights a security pattern called quarantine: agents that read untrusted public content should be kept away from high-privilege actions, while a separate agent handles actions that can change systems or data. That is a separation-of-duties model for autonomous AI work, and it will matter more if teams give agents broader access to repositories, tools and internal systems.
“A harness for every task: dynamic workflows in Claude Code”
— Thariq Shihipar and Sid Bidasaria, Anthropic
multi-agent AI coordination tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs And Guardrails Remain Open
Several details remain unclear from the source material. It does not provide a precise token-cost range, a benchmark showing reliability gains across task types or a public comparison against simpler single-agent runs. It also does not say how often workflows should be allowed to spawn large numbers of agents in production settings.
It is also not yet clear how teams will set default limits for budget, model choice, tool permissions and human review. The feature appears powerful for complex work, but its value will depend on how well teams define stop conditions and prevent unnecessary agent expansion.
AI task orchestration software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Developers Test Real Workflows
The next step is practical adoption. Teams using Claude Code will likely test dynamic workflows first on bounded jobs: reviewing a limited pull request, ranking a finite ticket set, checking citations in a report or running a pilot migration plan.
Anthropic points readers to Claude Code documentation for current usage details. The key question now is whether developers can turn dynamic workflows into repeatable patterns that improve reliability enough to justify the added token use and orchestration complexity.
complex AI workflow management
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Anthropic announce for Claude Code?
Anthropic described dynamic workflows, a Claude Code capability that lets Claude write a JavaScript orchestration harness and coordinate temporary subagents for a complex task.
Is this meant for every Claude Code task?
No. The source material says dynamic workflows use more tokens and are meant for complex, high-value work, not routine changes such as typo fixes.
What kinds of work could benefit?
Likely candidates include large refactors, security reviews, research reports, fact-checking, backlog triage and root-cause investigations where parallel work and independent review add value.
What is still unproven?
The source material does not give public benchmark results, exact cost ranges or standard guardrails for production use. Those details remain developing.
Source: Thorsten Meyer AI
