Claude can now run 1,000 AI agents in parallel with dynamic workflows

A bigger brain for big jobs

Anthropic has unveiled dynamic workflows for its Claude system, a feature that lets a single lead agent plan a massive task, fan it out to up to 1,000 sub-agents running in parallel, and then merge all the results together. The system is now generally available across Claude Code, the Claude API, Amazon Bedrock, Google's Vertex AI, and Microsoft Foundry.

Previously, Claude could delegate to sub-agents one conversation at a time, with every result stacking into its memory. Dynamic workflows move the orchestration logic into a JavaScript script that runs in the background while your session stays responsive. This means Claude no longer has to hold a thousand intermediate results in its context window — it just reads the final merged output.

How it actually works

To use the feature, users select the "multiagent_20261001" agent type and can start a workflow by simply asking Claude to create one or by turning on a setting called "ultracode," which sets the effort level to maximum and lets Claude decide when to spin up a workflow automatically.

The script that Claude writes can branch, loop, and check its own work. Independent agents can adversarially review each other's findings before anything surfaces to the user, which Anthropic says produces more trustworthy results than a single pass.

Common use cases include codebase-wide bug sweeps, security audits, framework migrations spanning thousands of files, and stress-testing a plan from several independent angles before committing to it.

The cost of scale

Anthropic warns that dynamic workflows can consume substantially more tokens than a typical Claude session and recommends starting with a small, scoped task to understand usage. The feature is available on all paid plans.

The approach has drawn skepticism from some quarters. A senior OpenAI engineer recently called agent swarms a massive waste of tokens, arguing that fanning work out to dozens of agents burns through compute for marginal gains.

Anthropic's own internal testing tells a different story. The company hid 70 bugs in a 116,000-line codebase and ran both a single agent and the dynamic workflow system. A single agent caught between 14 and 27 bugs per run, while the dynamic workflow consistently found 66 ¹. Whether those numbers hold across different types of work is still unclear, and Anthropic itself notes that users should test with their own workloads.

What this means

Dynamic workflows represent a shift in how Claude handles complexity. Instead of relying on one long conversation to coordinate every step, the system can now delegate at scale and verify results independently. For developers working with large, messy codebases, it could turn week-long debugging sessions into something that finishes in hours. The tradeoff is token cost, and it remains to be seen whether the accuracy gains justify the expense for all kinds of tasks.

Key facts

• Dynamic workflows let one lead agent orchestrate up to 1,000 sub-agents in parallel
• The orchestration runs as a background JavaScript script, keeping the user session responsive
• Anthropic's internal testing found 66 bugs in a large codebase versus 14 to 27 from a single agent
• The feature is generally available on all paid Claude plans and multiple cloud platforms
• Anthropic warns workflows can consume substantially more tokens than a normal session

Key facts

• Up to 1,000 sub-agents can run in parallel under a single dynamic workflow
• Anthropic's internal testing found 66 bugs vs. 14-27 from a single agent
• Works as a background JavaScript script so sessions stay responsive
• Generally available on all paid Claude plans, API, Bedrock, Vertex AI, and Foundry
• Anthropic warns workflows can burn substantially more tokens than typical sessions

Sources

1. Anthropic's Claude can now orchestrate up to 1,000 AI agents in parallel through dynamic workflows — the-decoder.com
2. Introducing dynamic workflows | Claude by Anthropic
3. Orchestrate subagents at scale with dynamic workflows - Claude Code Docs
4. Claude Managed Agents Dynamic Workflows: Up to 1,000 Parallel Agents — explainx.ai
5. Claude Code: Dynamic Workflows With 1000 Agents | innobu — innobu.com