Agent Goes Silent After Greeting

Your agent responds to the initial greeting but then stops responding to the caller’s inputs.

Symptoms

• The agent picks up the call and delivers the welcome message
• After the caller responds, the agent becomes unresponsive
• No error messages appear in the dashboard
• The call doesn’t disconnect, but the agent won’t speak or act further
• It happens consistently across multiple test calls

Common Causes

1. The model — a very small or older model may become unresponsive after the first turn
2. An exhausted BYO API key — a “bring your own” OpenAI key that has hit its rate or credit limit
3. A prompt that contradicts itself — overly complex or conflicting instructions cause the model to stall
4. A prompt that fills the context window — leaving no room for the conversation itself
5. Transcription not picking up the caller — the agent isn’t silent, it just never heard anything

How to Fix

Step 1: Check the model

1. Open the agent and go to Advanced Settings → AI Behavior & Model Configuration.
2. Check which model is selected.
3. If you’re on the smallest model available, switch to a mid-tier one and re-test. Very small models are cheap but noticeably less reliable at sustaining a multi-turn call with tools attached.
4. Publish the version and make a fresh test call.

The model list in the dashboard is the authoritative source for what’s available on your account.

Step 2: Verify your BYO API key

If you’ve connected your own provider key:

1. Go to Integrations and find the provider.
2. Check in that provider’s own dashboard that the account has credit, the key hasn’t expired, and you haven’t exceeded a rate limit.
3. If the key is exhausted, remove it to fall back to Dialora’s default integration.
4. Save and test again.

A key that runs out mid-call produces exactly this symptom: a normal greeting, then nothing.

Step 3: Simplify the prompt

1. Open Basic Settings and read the Agent Prompt.
2. Look for:
   • Contradictory instructions (“be helpful but never answer questions”)
   • Deeply nested conditions
   • Instructions that conflict with the agent’s tools
3. Cut it back to the core instructions, test, then reintroduce complexity a piece at a time.

Step 4: Check the token meter

The prompt editor shows live token usage against your model’s context window, split across the system prompt, your prompt, tool definitions, and the space reserved for the conversation.

If the meter is nearly full, the agent has no room to hold a conversation. Cut the prompt, remove tools the agent doesn’t need, and move reference material into a Knowledge Base rather than pasting it into the prompt.

Step 5: Rule out transcription

If the agent seems to hear nothing rather than fail to answer:

1. Go to Advanced Settings → Voice & Audio Settings and check the transcriber provider and model.
2. If the agent has multiple languages configured, make sure multilingual detection is on — otherwise the transcriber locks to the first language it hears and a caller who switches goes unheard.
3. Try a different transcriber provider and re-test.

Step 6: Confirm you published

Configuration changes live on your draft until you publish. If you fixed something and the live call behaves the same, check for a Draft badge on the agent and publish the version — see Agent Versions, Draft and Publish.