What was added

Matthias Bastian at The Decoder reported on October 9 that Anthropic is adding dynamic workflows to Claude Managed Agents. The managed-agent infrastructure was already there. Dynamic workflows are new. A lead agent creates a plan, distributes tasks to sub-agents, and merges the results when they are done. Up to 1,000 agents can run in parallel per execution.

[1]
An open blue folder with many blank cards fanned out and a red clip at the top
The gouache shows an open folder and a fan of blank cards. It stands for a lead agent splitting tasks across sub-agents, and it is not a news photograph., AI-generated illustration, not a news photograph

Bugs they hid themselves

Anthropic points to its own test. The team hid 70 bugs in a 116,000-line codebase. A single agent caught between 14 and 27 per run. The dynamic workflow consistently found 66. The Decoder writes that whether those gains hold on other kinds of tasks is not shown. The figures describe this test only. They are not a general find rate.

The report also says a senior OpenAI engineer recently called agent swarms a massive waste of tokens. That is a comment The Decoder cites. It is not a result of the bug test.

[1]

Cost, and what is still open

The report says these workflows can burn through a lot of tokens. Anthropic recommends starting small. The agent type used to turn the workflows on is called multiagent_20261001. The connection steps are in Anthropic’s documentation. This article does not turn that connection into a procedure.

The 116,000 lines and the 70 bugs are conditions Anthropic put into the test. The report does not include the full method, and it does not include a third-party reproduction. The figure of 1,000 is the parallel cap for one execution, not a daily volume already measured in a customer environment.

[1]