Artificial Intelligence
Anthropic adds 1,000-agent workflows to Claude
Anthropic added dynamic workflows to Claude Managed Agents, letting one agent coordinate sub-agents in parallel, while the company said internal testing showed higher bug-finding rates than a single agent.
- Anthropic added dynamic workflows to Claude Managed Agents, and the feature let a lead agent plan work, assign tasks to sub-agents and combine the results.
- The company said the system could run up to 1,000 agents in parallel in a single execution.
- Anthropic said it hid 70 bugs in a 116,000-line codebase and compared the results of one agent with the multi-agent workflow.
- In that test, a single agent found between 14 and 27 bugs per run, while the dynamic workflow found 66.
- Anthropic said users could turn the feature on by selecting the multiagent_20261001 agent type and advised starting small because it could use a lot of tokens.
Anthropic added dynamic workflows to Claude Managed Agents. The feature let one lead agent make a plan, divide work across sub-agents and bring the results together. The company said up to 1,000 agents could run at once in one execution.
The article said the cost case was unclear. It noted that a senior OpenAI engineer had described agent swarms as a waste of tokens, then pointed to Anthropic’s own test results instead.
In that test, Anthropic hid 70 bugs in a 116,000-line codebase. A single agent found between 14 and 27 bugs per run, while the dynamic workflow found 66.
Anthropic said users could switch the feature on by choosing the multiagent_20261001 agent type. It also said the workflow could use a lot of tokens and pointed users to documentation or a Claude Code command to begin.
Named in this story
Companies
- Anthropic
- added the dynamic-workflow feature to Claude Managed Agents
- OpenAI
- was cited through a senior engineer who criticised agent swarms
Products and systems
- Claude
- was the Anthropic system gaining the new orchestration ability
- Claude Managed Agents
- received the new dynamic workflows
- multiagent_20261001
- was the agent type users selected to enable the feature
- Claude Code
- was one way to start the feature with a command
How the source tells it
Measured and slightly sceptical, with the article qualifying Anthropic’s claim at each step.
- capability hype the headline centres a very large maximum-scale claim
- caution the piece repeatedly hedged on cost and whether the results would generalise