Skip to content

Agent issues

Troubleshooting

Most early agent problems come from prompt wording, missing conversation history, or tools firing at the wrong time. Narrow the issue first, then change one setting or instruction at a time.

Leave with one likely cause and one focused change to test.

Most first-launch problems are not infrastructure failures. They are usually prompt, memory, or tool-timing problems. Start with one test call, identify the exact failure, and change one thing at a time.

The agent does not remember previous answers

This is still the most common issue.

Symptoms:

  • the assistant asks for the same callback fact twice
  • the conversation feels like a series of disconnected turns
  • the assistant forgets the reason for the call after the caller already said it

Check both of these immediately:

  1. Conversation history limit is set in Technical
  2. {conversation_history} is present in the system prompt

If either one is missing, memory will not work the way you expect.

If the settings look correct, open Voice Debug Monitor and confirm that the prompt sent to the model really includes recent conversation history.

The agent never confirms the callback number

This usually means the prompt is too vague or the caller number is not being used clearly in the flow.

Fix it by checking these points:

  1. The prompt explicitly tells the assistant to ask whether {caller_phone_number} is the best number for the callback
  2. {caller_phone_number} is actually present in the prompt
  3. The assistant is not trying to ask for several details in one turn

If the assistant skips straight to the close, tighten the conversation-flow section of the prompt and retest.

The assistant asks too many questions at once

This is a prompt discipline problem.

If callers hear compound questions like "What is your name, company, and callback number?", update the prompt rules so they clearly say:

  • ask one question at a time
  • do not move on until the current fact is collected or clarified

The callback assistant works best when it advances step by step.

The Email Tool fires too early

This is one of the most important operational failures to catch.

Symptoms:

  • an email is sent before the callback number is confirmed
  • the handoff is missing the company name or reason for the call
  • the assistant ends the call before the callback request is complete

Check two things:

  1. The tool's Execution Description is strict enough
  2. The prompt makes it clear that the tool should run only after all callback facts are collected

Use Voice Debug Monitor to see the exact moment the model chose to trigger the tool.

The assistant talks too much

For a phone callback flow, long answers usually make the experience worse.

If responses are too long:

  • lower Max completion tokens into the 150 to 250 range
  • keep Temperature around 0.5
  • shorten the prompt rules so the assistant is told to keep answers short and professional

If the assistant still overexplains, the prompt is usually the bigger issue than the model setting.

The assistant sounds random or inconsistent

If the same prompt produces noticeably different behavior from call to call, reduce the room for improvisation:

  • keep Temperature between 0.3 and 0.5
  • simplify the prompt
  • remove optional branches that are not needed for the callback flow

For first-launch assistants, predictability is better than creativity.

The knowledge base is not helping

If you added a knowledge base and the agent still answers generically, check:

  1. the correct knowledge base is connected
  2. the prompt actually includes {context}
  3. the callback flow is still the main instruction, with knowledge-base content as support

Do not use a knowledge base to compensate for a weak callback prompt. Fix the base conversation first.

The assistant ignores important instructions

When the assistant repeatedly misses a core rule, the prompt is usually too soft or too crowded.

Improve it by:

  • putting the key rule higher in the prompt
  • shortening instructions that do not affect the callback flow
  • keeping the most important requirements explicit, such as confirming the callback number and asking one question at a time

If needed, use Important Note for short operational context, but do not turn it into a second full prompt.

A practical troubleshooting order

When a test call fails, use this order:

  1. Review the system prompt
  2. Review Conversation history limit
  3. Review tool execution descriptions
  4. Check the call in Voice Debug Monitor
  5. Check the final output in Call History

This sequence usually gets you to the root cause faster than changing model settings at random.