Wrong responses

Find out why an agent said the wrong thing, called the wrong tool or made up an answer, and the settings and tests that fix it.

When an agent gives a wrong answer, first find out which layer caused it. Open the call in Call History and work through the transcript:

Did the agent hear the right words? Compare the recording with the user lines. If not, see Transcription accuracy.
Did it have the right information? Check the dynamic variables in the Data tab, and the knowledge base chunks it retrieved in Knowledge and workflows.
Did it call the right tool with the right arguments? Check the tool call and tool result items in the transcript.
Did it follow the prompt? If the inputs were right, the prompt or model is the cause.

Knowledge base chunks retrieved during a call

Reproduce it#

Fix the problem where you can repeat it:

Test LLM in the builder chats with the agent in text, using the same prompt, tools and knowledge base. Paste the caller's words from the transcript to reproduce a turn.
The test panel's turn debugging lets you inspect each turn of a test conversation.
Simulation runs the agent against an AI-simulated caller, and test cases let you save a scenario and re-run it after every change.

See Test for these tools.

Causes and fixes#

Made-up answers#

Cause	Fix
The answer is not in the prompt or knowledge base	Add it to the knowledge base, or tell the agent what to say when it does not know
Retrieval returned the wrong chunks	Check Knowledge and workflows on the call. Improve the source documents, and tune the agent's knowledge base settings
The prompt invites guessing	Add a rule such as "If the information is not provided, say you will check and offer a callback"

AI QA scores calls for hallucination and knowledge base recall, so you can find these calls across many conversations.

Wrong tool or wrong arguments#

Cause	Fix
Vague tool descriptions	Say exactly when to call each tool and when not to
Arguments in the wrong format	Describe each parameter with an example. Keep `tool_call_strict_mode` on (the default) so arguments follow the schema
A tool result the model misreads	Return short, labelled results from your endpoint
A transfer to the wrong number	Use a static destination per tool, or list numbers clearly for inferred destinations; see Debug transfers

Ignoring the prompt#

Cause	Fix
Prompt too long or contradictory	Shorten it. Move reference material to a knowledge base. The builder's Tokens estimate shows the size
The task needs strict steps	Use a multi-prompt agent with states, or a conversation flow with nodes and edges
Model too small for the task	Try a larger model and compare in simulations; see Billing for prices
Randomness	`model_temperature` defaults to 0. It is not sent to GPT-5 family models
Early details forgotten on long calls	The workspace LLM token limit drops the oldest turns when the history is too long. Raise it, or use collected variables to keep key facts

Missing or wrong variables#

Unfilled {{variables}} reach the model as written. Set defaults with Default Dynamic Variables, and check the call's Dynamic variables and Collected variables in the Data tab. The precedence is agent defaults, then environment tag, then contact fields, then the request, then values collected during the call, then live overrides.

Off-topic or unsafe replies#

Turn on scope_boundaries in the Agent Handbook with a scope_description, so the agent declines unrelated requests and does not reveal its instructions.
Add guardrail output_topics (for example regulated_advice) with a fallback_message. Guardrails cost $0.01 per minute.

Compare versions#

When a change makes things worse, compare the published version with the draft and roll back if needed.