When One Agent Isn’t Enough: Claude Now Builds Its Own Team of Agents on the Fly

  • by

Full opportunity report: When One Agent Isn’t Enough: Claude Now Builds Its Own Team of Agents on the Fly on ThorstenMeyerAI.com — validation score, market gap, and execution plan.

TL;DR

Anthropic’s Claude has introduced a new feature enabling it to dynamically create and orchestrate multiple agents for complex tasks. This development allows for more effective handling of high-value, multi-faceted projects by simulating team-like collaboration within AI workflows. The capability is currently targeted at advanced, resource-intensive tasks and is not meant for simple corrections.

Anthropic’s Claude has introduced a new feature called dynamic workflows, which allows the AI to build and manage its own team of agents on the fly. This capability enhances Claude’s performance on complex, high-value tasks by enabling it to orchestrate multiple specialized sub-agents, rather than relying on a single, monolithic model. The development was announced as part of an ongoing evolution in AI orchestration, aimed at addressing the limitations of single-agent workflows in handling multi-faceted projects.

The new feature, dynamic workflows, is a technological advancement where Claude writes and executes small JavaScript programs to spawn, coordinate, and manage multiple sub-agents tailored to specific sub-tasks. Each sub-agent can operate with its own context window, focus, and model choice, enabling parallel processing and specialized judgment. This approach mimics a human team lead delegating work to specialists and reviewers, improving accuracy and thoroughness in complex workflows.

According to Anthropic, this system is particularly suited for high-stakes, complex tasks such as code refactoring, detailed research routines, or extensive fact-checking, where single-agent approaches tend to underperform due to issues like goal drift, self-bias, or incomplete work. The process involves various orchestration patterns, including classify-and-act, fan-out-and-synthesize, adversarial verification, and tournament-style comparisons, which collectively enhance reliability and depth of analysis.

Anthropic emphasizes that this capability is resource-intensive, using more tokens and computational power, and is not recommended for simple tasks like fixing typos. The feature is currently in deployment and can be triggered by specific prompts such as “ultracode,” signaling Claude to generate its own custom workflow for the task at hand.

At a glance
updateWhen: announced recently, with ongoing deploy…
The developmentClaude now autonomously constructs and manages its own team of specialized agents during complex workflows, marking a significant evolution in AI orchestration.

Claude Builds Its Own Team: Dynamic Workflows — Insights

AI Dispatch · Insights · 1 July 2026

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.

Why one agent grinding alone underdelivers
Agentic laziness
Declares done on partial work — 35 of 50 review items.
Self-preferential bias
Grades its own homework — likes what it already produced.
Goal drift
Loses the original objective across turns, especially after context is summarized.
These are the failure modes of one person doing a huge job alone. The cure is the manager’s: divide the work, give isolated briefs, and have someone independent check it.
The harness — an org chart Claude writes for one task
Orchestrator
Claude writes a JS harness on the fly
▼   fan out   ▼
Subagent
own context · model
Subagent
own worktree
Subagent
focused goal
Subagent
isolated
✕ adversarial verify
✕ adversarial verify
✕ adversarial verify
✕ adversarial verify
▼   barrier: wait for all   ▼
Synthesize
merge structured outputs
→ Result
one verified answer
Each subagent gets a clean context window and can run on a cheaper or smarter model — so no single overloaded context gets lazy, biased, or lost. Resumable if interrupted.
The six moves it composes
Classify-and-actroute by task type (switchboard)
Fan-out-and-synthesizeparallel agents → a barrier merges (map/reduce)
Adversarial verificationa separate agent attacks each result
Generate-and-filterbrainstorm wide, keep only survivors
Tournamentagents compete; pairwise judging > scoring
Loop-until-donespawn until a stop condition, not a fixed count
Where it earns its keep — often away from code
Big migrations & refactors
Deep research → cited report
Fact-check every claim
Rank 1,000 tickets by severity
Root-cause post-mortems (“why did sales drop?”)
Triage a backlog at scale
Design/naming by rubric
Model routing
One security pattern to memorize — quarantine: agents that read untrusted public content are barred from high-privilege actions; a separate agent does the acting. Separation of duties for autonomous agents.
The take

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.

Source: “A harness for every task: dynamic workflows in Claude Code,” Thariq Shihipar & Sid Bidasaria (Anthropic), Claude blog, 2 June 2026. Mechanics, patterns & use cases are Anthropic’s; the “org chart” framing is the author’s. A recent, still-evolving feature. Docs: code.claude.com/docs.
thorstenmeyerai.com

Implications for AI-Driven Complex Workflows

This development signifies a step toward more autonomous and collaborative AI systems capable of managing intricate projects without extensive human oversight. By enabling Claude to dynamically assemble and orchestrate its own team of specialized agents, organizations can potentially improve the quality, accuracy, and efficiency of high-stakes tasks such as research, software development, and detailed analysis. It also reduces the need for manual scripting of workflows, allowing for more flexible and adaptive AI applications.

However, the increased resource consumption and complexity mean that this feature is best suited for organizations with significant computational capacity and the need for advanced AI orchestration. It raises questions about the scalability, control, and transparency of AI systems that can self-organize and manage multiple agents independently.

Evolution of AI Orchestration and Multi-Agent Systems

Anthropic’s recent work on Claude reflects a broader trend in AI development toward multi-agent and dynamic workflow systems. Previous iterations focused on single-agent models performing isolated tasks, but recent innovations aim to mimic human team dynamics—delegating, reviewing, and iterating across multiple specialized agents. This approach builds upon earlier concepts like static multi-agent setups, but now with greater flexibility and automation, thanks to the ability of Claude to generate custom JavaScript workflows.

The concept of orchestrating multiple agents is not entirely new; however, the ability for an AI to write its own orchestration code on the fly marks a significant leap. This feature was developed alongside Claude Opus 4.8, which enhanced the model’s reasoning capabilities, allowing it to determine the appropriate orchestration pattern for each task.

While the technical details are still emerging, the approach reflects ongoing research into AI’s capacity to handle complex, multi-step processes more reliably than single-agent systems, which often suffer from goal drift, bias, and incomplete work.

“Claude’s ability to build its own team of agents represents a major advancement in autonomous AI workflows, enabling more reliable handling of complex tasks.”

— Thorsten Meyer, AI researcher

Unanswered Questions About Autonomous Agent Management

It is still unclear how widely this feature will be adopted in real-world applications or how it will perform outside controlled testing environments. The resource demands and complexity may limit its use to specific high-value scenarios. Additionally, the long-term implications for AI transparency, control, and safety remain to be fully understood, as the system’s ability to self-organize introduces new layers of autonomy.

Details about the limits of the current implementation, potential failure modes, and safeguards against unintended behavior are still emerging. Researchers are watching closely for how this capability evolves and whether it can be reliably controlled at scale.

Next Steps in Developing and Deploying Self-Organizing AI Workflows

Anthropic is expected to continue refining this feature, potentially expanding its capabilities and integrating more sophisticated orchestration patterns. Future updates may include more user controls, safety mechanisms, and broader deployment in enterprise environments. Monitoring and evaluation of real-world use cases will inform whether this approach can be scaled safely and effectively.

Organizations interested in this technology are likely to see pilot programs and case studies in the coming months, with further technical documentation and best practices following as the system matures.

Key Questions

What kinds of tasks is Claude’s dynamic workflow best suited for?

High-value, complex tasks such as detailed research, extensive code refactoring, multi-step fact-checking, and large-scale analysis are ideal candidates for this feature, where multiple specialized agents can improve accuracy and efficiency.

Does this mean Claude can now fully automate complex projects without human oversight?

Not entirely. While it can manage complex workflows, human oversight remains important, especially for setting goals, monitoring performance, and handling unexpected issues. The feature is designed to augment human effort, not replace it.

What are the resource implications of using dynamic workflows?

This approach uses more tokens and computational power, making it best suited for organizations with significant infrastructure. It is resource-intensive and not recommended for simple or low-stakes tasks.

Are there safety or control mechanisms in place for autonomous agent orchestration?

Details are still emerging, but Anthropic emphasizes that safeguards are crucial. The system includes verification and review steps, but the long-term safety implications are under ongoing study.

Source: ThorstenMeyerAI.com

Leave a Reply

Your email address will not be published.