What do the calls being sequential have to do with tokens? Do you just mean that the LLM has to think everytime they get a response (as opposed to being able to compose them)?
replyLLMs can use CLI interfaces to compose multiple tool calls, filter the outputs etc. instead of polluting their own context with a full response they know they won't care about. Command line access ends up being cleaner than the usual MCP-and-tool-calls workflow. It's not just Anthropic, the Moltbot folks found this to be the case too.
replyThat makes sense! The only flaw here imo is that sometimes that thinking is useful. Sub-agents for tool calls imo make a nice sort of middle ground where they can both be flexible and save context. Maybe we need some tool call composing feature, a la io_uring :)
reply