Depth first search via recursive dispatch is my current go-to strategy. This requires a fully serialized chain of operations. Not even tool calls are allowed to run in parallel. One agent and one action at a time. That's it.

This does feel really slow at first, but in practice it seems to converge on high quality solutions more quickly than breadth first techniques which leverage parallelism and more tokens. The solution itself might not be the most ideal, but we find it very quickly and with a great deal of consistency. This means we can iterate more rapidly and with more certainty.

Just because we can consume $100 of tokens in 60 seconds doesn't mean we should try to. I think trying to go fast is what's burning a lot of people out on AI. There's a ton of value in here if we can slow down and be a little bit more deliberate about it.