--- title: "Configure your voice agent's instructions and behavior" description: "Personality and Instructions become the agent's system prompt. Set the model it thinks with, how long it waits before replying, and let it search your website when your knowledge base comes up empty." --- > Documentation Index > Fetch the complete documentation index at: https://docs.speakai.co/llms.txt > Use this file to discover all available pages before exploring further. # Configure your voice agent's instructions and behavior Everything your voice agent says on a call comes from a handful of settings on its editor: who it is, what it's told to do, which model it thinks with, and how quickly it jumps in. Get these right and the agent sounds like it belongs in your business; get them wrong and callers notice immediately, because there's no visual polish to hide behind on a voice or video call, only what the agent says. ## Personality and Instructions are the agent's system prompt Open an agent and go to its **Personality** and **Instructions** fields. These aren't cosmetic descriptions. Every call starts with a system prompt built from them word for word: the agent is told "You are [agent name], an AI assistant with the following personality: [your Personality text]," and whatever you write in **Instructions** is appended underneath as its own section. The model reads that prompt before it hears a single word from the caller, so whatever you write here is what shapes every reply for the rest of the call. Use **Personality** for tone: warm and casual, brisk and professional, or anything in between. Use **Instructions** for what the agent should actually do: the questions to ask, the order to ask them in, how to handle a caller who goes off topic, when to escalate to a human. Keep the two separate rather than blending tone into the instructions text; it keeps both fields easier to revise later. Each field has its own **Generate** button. Select it to have AI rewrite that field, optionally with a short steering prompt of your own, without touching the rest of the agent's config. It's the fastest way to fix a Personality or Instructions field that's drifted from how the agent actually behaves, rather than rewriting it by hand. ## Choosing a model The **Model** picker controls which LLM the agent thinks with, split into two groups: - **Standard** models listen to the caller, think, then speak. The picker shows each one's provider and its per-million-token input and output pricing, so you can weigh cost against quality. - **Live (speech to speech)** models hear the caller and speak directly, without a separate listen-then-speak step. The picker shows the provider, a per-minute rate, and how many of that model's own voices are available. Switching to a Live model changes what else on the agent applies. Response pace, pronunciation rules and keyword boosting stop applying, and scripted lines like the welcome message are spoken close to, but not exactly, as written, because the model is generating speech directly rather than reading a script through a separate text-to-speech voice. A Live model also speaks with its own voice catalog: pick one on the agent's Voice & Avatar tab, since the regular text-to-speech voices aren't available with a Live model. A model change applies to the agent's next call, never to a call already in progress. ## Response pace **Response pace** sets how long the agent waits after the caller stops talking before it replies. It only applies to Standard models; the setting is unavailable when the agent is on a Live model, since Live models generate their own timing. - **Snappy**: replies as soon as you stop talking. Best for quick, simple calls. - **Balanced**: works well for most conversations, and is the default. - **Patient**: waits longer when callers pause to think. Good for support and intake calls. - **Very patient**: waits longest. Best for callers reading out numbers or other details where a pause mid-sentence is common. Like the model, a response pace change applies to the agent's next call, not one already underway. ## Creativity level Every agent also carries a **creativity level**, a number from 0 to 1 that defaults to 0.5. When Speak AI generates or refines an agent from a prompt, it sets this automatically for the kind of agent you describe: lower, around 0.2 to 0.3, for support and healthcare agents that should stick closely to the facts, and higher, around 0.6 to 0.7, for coaching or education agents that benefit from more varied phrasing. It isn't a field you set directly in the agent editor today; it travels with the rest of the agent's config through the API and through AI generation. ## Give the agent a website to search Set **Website URL** on the agent's Details section, and the agent can search that site during a call to answer questions your knowledge base doesn't cover. This only takes effect once the agent also has at least one knowledge base collection attached (on the agent's Knowledge Base tab) and web search turned on for the agent: with both in place, the agent is instructed to search its knowledge base first, and only fall back to searching the website when the knowledge base returns nothing or a low-confidence answer for an in-scope question. It won't use website search for off-topic requests or general web lookups, and it calls it at most once per question, speaking a short acknowledgment first so the call doesn't go silent while it searches. Today, turning web search on for an agent isn't exposed as a switch in the agent editor; it's set through the Speak AI API's `enableWebSearch` field alongside `websiteUrl`. If you want an agent to search its own site, set both through the API, or ask Speak AI support to enable it for the agent. ## Related pages - [Create and manage Speak AI voice agents](/api/voice-agents/agents/) for the full `personality`, `instructions`, `llm`, `websiteUrl` and `enableWebSearch` field reference, and how to set them through the API. Source: https://docs.speakai.co/help/voice-agents/instructions-and-behavior/index.mdx