I usually have a high-level agent drafting plans, then spawning sub-agents to build parts of the plan, other sub-agents to check the work of those agents. Then the PRs have different code review agents looking at the code, on top of the agents for people on my team also reviewing my code.
And I'm probably on the more amateurish end of AI development. As I learned these techniques from YouTube.
When you hear about the work Anthropic researchers are doing, they seem to be using hundreds or thousands of agents to solve problems.
> When you hear about the work Anthropic researchers are doing, they seem to be using hundreds or thousands of agents to solve problems.
There's a reason they're Anthropic researchers. The value proposition tanks when you actually have to foot the bill for the tokens you use.
Subagent flows can improve overall performance, but it generally also comes with disproportionately high token use. That gets worse with parallel subagents that are often doing the same things at the same time and cannot share a cache.
They are doing that because they are running into hardware limits with single agents. If one agent made correct decisions and explored spaces efficiently, they would not need hundreds of thousands of agents. It's is the luxury of great wealth that they can produce a large number of possibly defective samples and then select the ones that are not defective.
And I'm probably on the more amateurish end of AI development. As I learned these techniques from YouTube.
When you hear about the work Anthropic researchers are doing, they seem to be using hundreds or thousands of agents to solve problems.