Voiceflow Google Gemini Agent crashing
Hello Guys, so I am building this build here using the agent feature with gemini 2.5 pro because it is the most fluent AI model in mongolian but everytime the chat leads to the agent it just ends conversation. How can I fix thi? This is urgent
22 Replies
Quick checklist to get your Agent running again ⤵️
1. Verify the model ID
– In the Agent settings make sure the dropdown actually says “Gemini 1.5 Pro” (that’s the only Google Pro model Voiceflow exposes right now). If you typed-in or pasted a custom model name such as gemini-2.5-pro, Google returns an error and the runtime just ends the turn.
2. Look at the debug trace
– Open the Test panel ▶️ switch to Debug.
– If you see something like
• Wait a few minutes – Google quota spikes are usually temporary. 5. Make sure the Agent has something to say – If there’s no “Assistant Introduction” or first Prompt, the chat will look like it died even though the model answered. Turn the setting on or add an explicit “Hi, how can I help?” Text step before you hand over control to the Agent. After going through the steps above the Agent should stay alive even when Gemini misbehaves, or at least give you a clear error in the Debug panel instead of silently ending the chat. Docs for reference: https://docs.voiceflow.com → LLM Fallback section
– If you see something like
provider_request_failed or quota_exceeded coming from google, gemini-pro-1.5, the request never reached the rest of your flow – so the conversation simply stops.
3. Add a fallback model (recommended)
Voiceflow can automatically fall back to another LLM when Google errors out:
Settings → LLM Fallback
• choose GPT-3.5-Turbo (or any model you trust) in the “Fallback for Google” column.
After you save, run the Test panel again – you’ll notice a debug line similar to:
"google, gemini-pro-1.5" on the fallback chain suffered an outage, falling back to "openai, gpt-3.5-turbo"
and the conversation will continue instead of ending.
4. If you need Gemini specifically
• Try switching to Gemini 1.5 Flash (lighter but same language coverage).• Wait a few minutes – Google quota spikes are usually temporary. 5. Make sure the Agent has something to say – If there’s no “Assistant Introduction” or first Prompt, the chat will look like it died even though the model answered. Turn the setting on or add an explicit “Hi, how can I help?” Text step before you hand over control to the Agent. After going through the steps above the Agent should stay alive even when Gemini misbehaves, or at least give you a clear error in the Debug panel instead of silently ending the chat. Docs for reference: https://docs.voiceflow.com → LLM Fallback section
like-gold•2mo ago
Are you using a capture step before the agent to get the user's message?
button step
I don't see the Gemini 1.5 pro version here. There is only Gemini 2.5 Pro, Flash and 2.0
like-gold•2mo ago
is exit every conversational turn in the agent step toggled on?
all good questions. can you screenshot your agent settings?
Yeah, here it is.
Happened once again and this keeps happening. The conversation ultimately just ends when Gemini is used in the Agent block. The prompt block is good but after just a few tries it cuts off mid response and moves to the next node without finishing it. I need this to get fixed ASAP. Thank you.
It doesn't even look like the Agent step is starting
exactly. how do I fix this?
@NiKo | Voiceflow @Braden Any thoughts
@(0))) Try dragging a new Agent step, select the same prompt and connect it to your flow in place of the old one. Did that work?
Mm nope, same outcome. It either stops the conversation right away or responds with an unfinished reply. Like it stops mid word.
What do you have in your logs when you run a test?
Also, could you check if you face the same issue when using Claude 4 Sonnet or GPT 4.1?
Yes, the same error occurs with GPT 4, Claude 4.
it says execution failed when the scenario is chat’s end.
try this, add the agent right after the star block and see if the agent runs at all.
it just ends right away. right as I hit the test button
like-gold•2mo ago
@W. Williams (SFT) - could the mixed language be a factor here? Prompting in English + Mongolian?
@(0))) - could you try prompting the agent entirely in English and see if it works? You can add a line (in English) explaining it to respond in Mongolian
(might be a good idea to save your current prompt in a doc first so you don't lose it)
okay, let me try that. Thank you.
Tried it, same outcome. The conversation just ends.
yeah seems like agent failing... you're using a gemini model for the agent? @(0)))
like-gold•2mo ago
he said he's primarily using gemini but tried claude 4 and gpt as well - got the same outcome
My primary choice of LLM model is gemini because it is the most fluent in my client’s requested language. Is there a way to fix this within 2 days? If not can I get my voiceflow pro plan refunded? or atleast get another month of pro plan? Thank you for trying to resolve this with me. Highly appreciate it.