TL;DR
Anthropic has described a Claude Code feature called dynamic workflows, which lets Claude write a task-specific JavaScript harness and coordinate temporary subagents during one job. The feature is aimed at complex, high-value work, but its exact cost, access, and independent performance gains remain unclear.
Anthropic’s Claude Code has added a feature called dynamic workflows that lets Claude write a task-specific JavaScript harness and coordinate temporary subagents inside one job, according to source material citing Anthropic’s June 2, 2026 Claude blog post. The change matters because it pushes advanced coding agents beyond a single-worker model and toward managed parallel work for larger tasks.
The feature works by having Claude write a small JavaScript program that acts as an orchestration layer around the model. That harness can spawn separate subagents, give each one a focused brief, wait for their outputs, and merge the results into a final answer or code change, according to the source material.
Each subagent can have its own context window, focused goal, and in some cases its own worktree or model choice. Anthropic’s stated patterns include fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament-style judging, classify-and-act routing, and loop-until-done workflows that continue until a stop condition is met rather than a fixed number of steps.
Anthropic’s caveat is part of the news: dynamic workflows use meaningfully more tokens and are meant for complex, high-value tasks, not routine edits. The source material points to use cases such as large code migrations, deep research reports, claim-by-claim fact-checking, ticket ranking, post-mortem analysis, backlog triage, design selection by rubric, and security review.
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.
Bigger Tasks Get Parallel Agents
The feature addresses a known problem in agentic work: one model instance can lose track of a long task, favor its own previous answer, or mark work complete before the job is actually done. The source material describes these failure modes as agentic laziness, self-preferential bias, and goal drift.
Dynamic workflows try to reduce those risks by splitting work across isolated subagents and using separate agents to review or challenge results. For developers and technical teams, that could make Claude Code more useful on large, parallel tasks where a single context window becomes overloaded.
The trade-off is cost and control. More subagents can mean more tokens, more moving parts, and a greater need for clear limits. The source material says one security pattern to remember is quarantine: agents that read untrusted public content should be blocked from high-privilege actions, while a separate agent performs those actions.
Coding with AI For Dummies (For Dummies: Learning Made Easy)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Claude Code’s Third Agent Pattern
Thorsten Meyer AI frames dynamic workflows as the third part of a loose Claude Code arc: skills package organizational knowledge, loops decide how far to delegate over time, and dynamic workflows decide how to divide work inside a single task. That framing is the author’s, while the mechanics and patterns are attributed to Anthropic.
The underlying Anthropic source is identified as “A harness for every task: dynamic workflows in Claude Code” by Thariq Shihipar and Sid Bidasaria, published on the Claude blog on June 2, 2026. The source material also points readers to https://code.claude.com/docs for Claude Code documentation.
The feature fits a wider move in AI products from one-shot chat responses toward agent orchestration, where software coordinates multiple model calls, tools, files, and verification steps. What is new here is that Claude Code can generate the task-specific harness itself rather than relying only on a fixed workflow chosen in advance.
“Meaningfully more tokens”
— Anthropic caveat summarized by Thorsten Meyer AI
Hierarchical Sub-Agent Orchestration in CrewAI and LangGraph: Optimizing task delegation and state-persistence to reduce token overhead and prevent agentic loops
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs and Results Remain Unproven
Several details are still not clear from the source material. It does not give exact pricing effects, a public availability list, rollout limits, or independent benchmark results showing how much dynamic workflows improve outcomes compared with a well-run single agent.
It is also unclear how often Claude will choose the right workflow, how users can audit a generated harness, and how teams should set token budgets, stop conditions, or approval gates when a task could spawn many subagents. The source describes the feature as recent and still developing, so operational rules may change as users test it.
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.
Teams Move Toward Bounded Pilots
The next step for users is likely controlled testing on bounded, high-value tasks: a migration slice, a research report with citations, a ticket-ranking batch, or a security-review pilot. Teams will need to compare output quality, token use, review burden, and failure modes against single-agent Claude Code runs.
Further updates are expected through Claude Code documentation and Anthropic’s Claude blog. Until more results are public, the clearest guidance from the source material is practical: use dynamic workflows when work is big, parallel, adversarial, or judgment-heavy, and avoid using a multi-agent setup for simple edits.
LS Gen 3 ECM Engine Programming Control Module Reader Diagnostic PCM Bench Harness w/Power Supply Fits for GM Chevy All Blue/Red & Blue/Green 1999-2007, OBD2 Port, 18 Gauge, 12 Volt
Gen 3 LS ECM Programming Bench Harness Application – This bench-top harness that plugs into a standard wall…
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 Code can write to spawn, coordinate, verify, and merge work from temporary subagents during one complex task.
Is this just opening several Claude chats at once?
No. The source material describes a JavaScript harness that Claude writes and runs, with subagents given separate briefs, isolated context, and structured outputs that are merged at the end.
Who should use the feature?
Anthropic’s caveat is that the feature is for complex, high-value work. Good candidates include large refactors, migrations, research reports, claim verification, backlog triage, and security reviews.
Does it make Claude more reliable?
The source material says separate reviewers and adversarial checks can reduce single-agent failure modes. But independent performance data and real-world cost comparisons are still not provided in the material.
What is the main risk for teams?
The main risks are higher token use, overdelegation, and weak controls around agents that read untrusted content. The source recommends quarantine and separation of duties for higher-risk workflows.
Source: Thorsten Meyer AI
