Regarding this point that I thought many people knew, but discovered during company lectures that everyone actually doesn't know about agent development key techniques: tool-call, whether function_calling or MCP calls, are divided into explicit and implicit types. 1️⃣ Implicit Calling: Modern Automated Tool Execution Recent LLMs have built-in automatic tool calling capabilities. For example, (you can control this feature through the automatic_function_calling flag). Since official documentation and examples mainly focus on implicit calls, if you have been in MCP development in the last two months, you have likely been using the implicit mode. In the implicit mode, the entire tool calling process is a black box for developers. The LLM automatically determines whether tools need to be called, automatically selects appropriate tools and parameters, and automatically executes tools and integrates results. Developers only need to provide a list of tools and then directly obtain the final natural language answer. Implicit is the newer technology and the future trend. As large model capabilities evolve, this should become mainstream. Smart colleagues probably guessed it - after the praise, the turning point comes. Yes, implicit calls have temporary, hard-to-avoid drawbacks.

3️⃣Explicit call: manual parsing mode with precise control In contrast, the explicit call mode requires developers to manually handle each step of the tool call. You need to tell LLM how to format the tool call request in the system prompt, then manually parse the tool call instructions in the LLM answer, execute the corresponding tool, and finally integrate the results into the final answer. This is the ancient method of calling MCP (although it has only been traditional for half a year), and implicit only handles this part by LLM. However, in fact, the explicit call used in actual development is not completely ancient.

4️⃣ Semi-automatic backfill display call The method name is randomly chosen by me. The specific method is very simple. You need to give the LLM tool list to the MCP Client for the user's query, and ask it not to call the tool itself, but to return the name of the tool to be called. Next, you need to let the Client manually call the MCP Server. Since this part of the data is likely to be the formatted JSON data returned by the ordinary API, you can encapsulate this and the LLM response and return it. This method is not so widely applicable because it requires multiple rounds of LLM dialogue, which is still much more troublesome than a single dialogue. At the same time, your MCP HOST is better not to be a completely universal chat client that manages everything. This is more suitable for fixed services. But the advantage is that with raw data, you can process the data completely freely, and the data processing and UI display codes can directly reuse the parts of the previous service. And this method is compatible with LLM that does not have automatic tool calls. The new service led by me in our company adopts this method, and the response is good. Are there any other methods? I would like to hear your experience: @dotey @chaowxyz @hylarucoder @yihong0618 @pangyusio

Substack version: open.substack.com/pub/web3rove...…

Quaily: quaily.com/cryptonerdcn/p/disc...… @lyricwai @QuailyQuaily

Some responses to various misunderstandings: