Certification: Agent Design & Tools
This section tests your understanding of how to design effective voice agents — writing instructions, building tools, orchestrating multi-agent systems, and managing session lifecycle.
Agent instructions function_tool Multi-agent handoff Session management RunContext Userdata
Test your knowledge
Question 1 of 5
A client's agent gives long, rambling responses that feel unnatural in voice. What is wrong with the agent instructions and how do you fix it?
A The instructions lack voice-specific constraints. Add explicit guidance: 'Keep responses to 1-2 sentences. Never use markdown, bullet points, or emojis. Speak conversationally — you are a voice agent, not a chatbot.' Voice agents need instructions optimized for spoken delivery, not text output. B Set max_tokens on the LLM to 50-80 tokens per response. This hard limit forces the model to produce concise output regardless of instructions. Combine with a temperature of 0.2 to reduce verbosity. The instructions themselves are less important than the model parameters for controlling response length. C Enable the TTS provider's sentence chunking mode, which breaks long responses into individual sentences and only synthesizes the first 1-2 sentences. The remaining text is discarded before it reaches the audio pipeline. This approach works with any LLM instructions and requires no prompt engineering. D Configure the pipeline's response_filter middleware to truncate LLM output at 2 sentences before passing it to TTS. This is more reliable than instructions because LLMs do not consistently follow length constraints. The filter uses punctuation detection to find natural sentence boundaries.