Hi team, Today suddenly all our production agents have started to forget to call tools. Instead of calling the tool to look up slots or book appointments, it outputs a filler like "let me check that for you" and doesn't call the tool. This is happening across both our production and staging orgs on weekly channels. - We're using OpenAI GPT5.2 and it doesn't appear to be any degradations. - We've also determined no regressions in prompt Could we please get assistance here? SAmple CAll IDs: - 019c9cc8-ff2a-7ffe-9d1f-2c4ad993fa9b - 019c9ca3-9e88-7228-a38e-b3b1ee550ec6

Immediate Action: Go to your Vapi dashboard and navigate to

Observe > Call Logs

. Select one of your sample Call IDs and review the transcript and tool execution section to see if tool calls were attempted or if any errors are logged. This will help you quickly identify if the issue is with tool triggering, configuration, or model output. Example path:

Dashboard → Observe → Call Logs → [Select Call ID] → Tool Execution

For more details, see the [Debugging voice agents documentation](https://docs.vapi.ai/debugging). Source: - [Debugging voice agents documentation](https://docs.vapi.ai/debugging)

use gpt 4.1

Thanks - that doesn't help

any updates? I am facing the same issue.

I had to build in reinforcements to the prompt to force immediate tool calls. Also added in time-out hooks so if it happens and theres silence, it will trigger a human response which will kick start tool call

thanks.

Hi, I checked both calls and found the issue. Your agent is choosing to ask clarifying questions instead of calling tools when the user input is vague. First, tighten your prompt with stronger tool-calling instructions: Add this at the top of your system prompt:

CRITICAL: When a patient mentions ANY appointment-related keyword 
("cleaning", "checkup", "appointment"), call the query tool IMMEDIATELY. 
Do NOT ask clarifying questions first. Call the tool, then use the results 
to guide the conversation.

Then, bump your

maxTokens

from 250 to 500 in your assistant settings to give the model more room for tool calls. Test a few calls after these changes and let me know if you're still seeing issues! Best regards, Vapi Support

Hey @Shaunak That's interesting, we have a few other tools this is happening to and it's only happened on that day. We've added reinforcements and it seems better now, but curious to know if any known issues on that day. Bumping maxTokens up is interesting to me, my understanding is that we'd also have to prompt in "You are not allowed to speak more than x words per turn" to reduce yapping too.

Why this happened all of the sudden? Did VAPI issued new updates?

Hi, Glad the reinforcements helped. To answer your question - we don't have any known platform issues logged for that specific day, so it may have just been the model behaving inconsistently with vague inputs. On the

maxTokens

point, increasing it gives the model more room to complete tool calls, but if you're worried about verbosity you can balance it with a prompt instruction like "keep responses under X words per turn." Best, Vapi Support