I have been experimenting with optimizing my local-LLM agentic AI workflow. Currently I have Gemma 4 E4B (other models work well too, but this one seems reliable and fast) as the LLM.
In testing my system, I found that a simple query, like "summarize the Kimi-K3 blog post" was slower than necessary because the agent had to call multiple tools to resolve the query. It had to find the date (with find_blog_date
or similar) and then read_blog_post
.
I decided to try to reduce and combine tool calls—the read_blog_post
MCP tool call could internally perform find_blog_date
. The instructions on the tools would be updated to note the new calling pattern. So I could resolve a query with 1 tool call instead of 2 tool calls (and this also reduces thinking and token use).
Basically, combining tool calls—such that a single tool call can resolve its own arguments—and not relying on the agent seems beneficial to performance. This makes the agentic AI workflow faster than before, reduces expensive GPU calls, and leads to greater user satisfaction.