Comment by HarHarVeryFunny

For anyone, like me, who is/was confused by the patent diagram and labeling, the initial arrows labelled "tool call execution request" (client->server) and "tool calls" (server->LLM) seem to be confusingly labelled. This isn't a tool call being sent from client, via server, to LLM (which is not a thing - tool calls go in the opposite direction), but rather the user/agent's prompt, which as always includes tool "declarations", being sent from agent, via server, to the LLM.

The big picture here is that overall tool call flow works as normal, but rather than the LLM just generating tool calls, it is instead generating a "code block" that may include multiple tool calls/invocations. The server (between LLM and client) receives this code block and will execute it, but first has to execute any contained tool calls, some which it may do locally, and some via calls back to the client/agent as normal. Once all tools calls and the containing code block have been executed by the server (in a sandbox), then the final result is sent back to the client.