Goal
Create an AI agent that can answer common support questions and escalate unresolved cases to a human operator.Before you start
- Connect an AI provider first if you are using BYOK.
- Decide which channel the agent will handle first: voice, messaging, or both.
- Prepare any files, examples, or repository knowledge you want the agent to use.
Step 1: create the agent
InAI agents:
- name the agent clearly by function or queue
- choose the realtime provider and model
- set the voice, tone, and prompt baseline
Step 2: define the support behavior
Review:- how the agent should greet and guide people
- which product questions it should answer directly
- when it should request operator attention
- whether it should send follow-up messages or email
Step 3: add knowledge and tools
Use:- uploaded files for internal reference material
- GitHub repository bindings for issue and product context
- Discord examples when you want the agent to mirror approved support answers
Step 4: test before publishing
Use test sessions to validate:- the prompt quality
- the escalation behavior
- the tool choices
- the tone and pacing
Example workflow
- a customer asks a product question
- the agent answers from internal knowledge and repository context
- if the issue looks like a bug or needs account-specific help, the agent escalates to an operator or issue workflow
What good looks like
- the agent answers common questions consistently
- it uses only the tools it needs
- it escalates before confidence drops too far
- production publishing happens only after successful tests
Troubleshooting
- If the agent is too broad, reduce tools and tighten the prompt.
- If it escalates too late, make the escalation criteria more explicit.
- If answers are stale, refresh connected files or repository bindings.