Claude Dynamic Workflows Run Up to 1,000 Agents at Once
Anthropic has added dynamic workflows to Claude Managed Agents, its hosted platform for running AI agents. The managed agent infrastructure itself is not new. What changes is that Claude can now coordinate a whole team of agents inside a single run, with up to 1,000 of them working in parallel per execution.
How dynamic workflows work
The setup follows a manager-and-workers pattern. A lead agent looks at the task and writes a plan. It then splits the work into pieces and hands each piece to a sub-agent. When the sub-agents finish, the lead agent collects their output and merges it into one result.
That is the core idea. One agent does not work through a long job step by step. Instead, the job is broken apart and many agents handle it at the same time. The limit of 1,000 parallel agents per execution sets how far that split can go.
The bug-hunting test
Anthropic supports the feature with numbers from its own testing. The team planted 70 bugs in a codebase of 116,000 lines and set agents loose on it.
- A single agent found between 14 and 27 bugs per run.
- The dynamic workflow found 66, and it did so consistently.
On that task, the gap is large. At its best, the single agent caught a little under 40 percent of the hidden bugs. The multi-agent setup came close to finding all of them. The spread matters too: the single agent's results ranged from 14 to 27, while the workflow's result held steady.
One test only shows so much. This is a single benchmark, built and run by the company selling the product. It is still unclear whether the same gains appear on other kinds of work beyond finding bugs in code. The practical advice is the usual one: try it on your own workloads before drawing conclusions.
The cost question
Not everyone thinks swarms of agents are worth it. A senior OpenAI engineer recently called agent swarms a massive waste of tokens. Tokens are the units of text that language models read and produce, and they are also what customers pay for.
Anthropic does not hide this side. The company says dynamic workflows can use "a lot of tokens" and recommends that teams start small. The math explains why. A run with hundreds of sub-agents, each reading part of a codebase and reporting back, adds up quickly. Teams that have been watching their spend since Anthropic cut prices on Claude Haiku 5.5 will want to check how a large multi-agent run changes their bill.
Getting started
To turn on dynamic workflows, developers select the agent type "multiagent_20261001". Anthropic offers two ways in. Developers can follow the documentation, or they can run the command "/claude-api managed-agents-onboard" inside Claude Code, the company's coding agent.
Our Take
This suggests Anthropic sees orchestration, and not only model quality, as a place to compete. Anthropic now offers a lead agent with sub-agents as a managed feature. Teams that want this setup no longer have to build and maintain that coordination layer themselves. For developers already reshaping Claude Code to fit their own work, it adds another building block.
The open issue is cost against benefit. The bug test is striking, but it covers one task chosen by the vendor. The OpenAI engineer's criticism shows the industry has not settled whether swarms are efficient or just expensive. It is worth watching whether independent users report similar gains on other tasks. Another question is whether agent platforms add firmer spending controls, like the token caps Cohere introduced on its own agent platform. A thousand agents is a big number. Whether it is a useful one will depend on the job.