How the AI Does Things Instead of Just Talking

Questions this answersQuestions this answersfaq

can the AI change something in our system by mistake?can the AI change something in our system by mistake?what actually executes a tool call?

Not by itself. It can only ask. Your code is what runs, and your code can say no. If something gets changed that should not have been, that is your code agreeing to it, not the AI reaching in. This is good news, because it means the rule that stops it lives somewhere you control and can test.Not by itself. It can only ask. Your code is what runs, and your code can say no. If something gets changed that should not have been, that is your code agreeing to it, not the AI reaching in. This is good news, because it means the rule that stops it lives somewhere you control and can test.Your code does. A tool call is a structured request naming a tool and its arguments; nothing runs until your handler runs it. That makes the allowlist, the argument validation and the write gate ordinary code with ordinary tests. A rule written only in the system prompt is a preference the model can talk itself out of; the same rule as an if-statement cannot be argued with.

why did our AI stop doing several things at once?why did our AI stop doing several things at once?why did parallel tool calls stop?

Almost always because of how the answers were sent back. If the AI asks for three things and your code replies three separate times, it learns that asking for three at once did not work, and it goes back to asking one at a time. Run them together and reply once with all three results.Almost always because of how the answers were sent back. If the AI asks for three things and your code replies three separate times, it learns that asking for three at once did not work, and it goes back to asking one at a time. Run them together and reply once with all three results.Results returned across multiple user messages instead of one. The model reads that as the parallel request having failed and reverts to serial calls, turning one round trip into several. Execute concurrently, then return every result in a single user turn, including failures marked as errors. A dropped result is worse: the model re-asks for the same call.