Level 3 complete
Level 3 complete! You’ve reached the highest level of PolyAcademy. You can structure complex projects, design reliable flows, and polish agent speech to production quality.
You covered:
- Code organization: utility files, imports, and the
conv.functionspattern - Flow fundamentals: steps, persistent prompts, transitions, and exit behavior
- Flow patterns: collection, validation, verification, and deterministic vs LLM-led control
- Writing natural agent speech: brevity, tone, avoiding repetitive structure
- Voice polish: filler, turn-taking, personalization, and voice quality
You can now answer without guessing:
- How should I structure code so it stays maintainable?
- When should I use a flow vs a function vs a managed topic?
- How do I enforce strict retry limits while keeping speech natural?
- What makes agent speech sound human rather than generated?
Level 3 recap quiz
Check your understanding
You need the same phone number validation in 5 different flow functions. What's the best approach?
- Copy the validation regex into each function
- Define it once in a utility file and import it everywhere
- Use conv.functions to share the validation
- Add the validation to the Behavior field
Answer: Define it once in a utility file and import it everywhere
Define the validation in a utility file and import it with standard Python syntax. Update it in one place, fix it everywhere – this is the core benefit of modular code.
A flow collects a phone number in step 1, but users sometimes give their name first. What should happen?
- Reject the name and repeat the phone number question
- Accept the name, save it, and still ask for the phone number – handle information in any order
- Skip the phone number step since the user volunteered different info
- Exit the flow and start over
Answer: Accept the name, save it, and still ask for the phone number – handle information in any order
Design principle 2 (listen robustly): handle information in any order. If the user gives their name unprompted, save it and continue collecting the phone number. Don’t discard useful information.
A user fails identity verification three times. The flow should transfer to a live agent. How do you implement the transfer message?
- Hard-code a fixed phrase like ‘Transferring you now’
- Let the LLM generate the message freely
- Pass the utterance as a function argument so the LLM generates context-appropriate phrasing with deterministic execution
- Use a Response Control to replace the agent’s message
Answer: Pass the utterance as a function argument so the LLM generates context-appropriate phrasing with deterministic execution
Passing the utterance as a function argument gives you LLM-generated phrasing (context-aware) with deterministic execution (the handoff always happens). Best of both worlds.
The agent says 'In order to look up your account, I'll need your phone number. Could you please provide me with that?' – on every single turn. What are two things wrong?
- The response is too short and needs more detail
- The explanatory preamble is unnecessary, and the phrasing is too formal – simplify to ‘What’s your phone number?’
- The agent should ask for email instead
- The agent should use DTMF input for phone numbers
Answer: The explanatory preamble is unnecessary, and the phrasing is too formal – simplify to ‘What’s your phone number?’
Two issues: the explanatory statement (‘In order to…’) is filler the user can infer from context, and ‘Could you please provide me with’ is overly formal. Combine into a single natural question.
What’s next?
You’ve completed all three levels. From here, the reference docs and maintain guides are your ongoing companions: