Intelligence

Intelligence controls the language model behind an agent’s responses. It is the third of the four pills at the top of the agent editor, after Language and Transcriber. Unlike the Transcriber pill, nothing here is filtered by language: every model in the catalogue handles every language Vocily supports.

​

Choose a model

Open an agent, open the Intelligence pill, and choose a model. The picker is served from Vocily’s live catalogue and lists the model’s provider and its per-minute rate - there are no marketing descriptors. The list is ordered cheapest first. The model is used by the live voice pipeline and by the agent’s browser and test experiences. Choices are validated server-side, so a model shown as coming soon cannot be activated by editing the request or configuration directly.

​

Tune response parameters

Temperature and Max Tokens live under the pill’s gear.

​

Temperature

Temperature controls how much variation the model can use when producing a response. The slider runs from 0 to 2 and starts new agents at 0.4.

• Lower values make the agent more focused and repeatable.
• Higher values allow more creative variation, but can make a tightly scripted agent less consistent.

For customer support, booking, qualification, or policy flows, start low and raise the value only when the agent sounds too rigid. Temperature does not replace clear instructions in Agent Core.

​

Max Tokens

Max Tokens sets the maximum response length for one model turn. It accepts 64 to 4096 and defaults to 256. Keep this limit aligned with the channel:

• Phone agents usually work best with short replies and a modest limit.
• An agent that needs to explain several options may need a larger limit.
• A larger limit does not require the agent to use all of it; your prompt should still say how concise the response should be.

When tools are enabled, allow enough response space for the agent to complete an action. This does not mean the agent needs to speak a long answer to the caller.

​

A practical starting point

1. Start on the default - GPT-5.4 Nano, Temperature 0.4, Max Tokens 256.
2. Make the response style in Agent Core explicit: keep phone replies short, ask one question at a time, and do not invent facts.
3. Test the same path in Chat, Talk, or Call.
4. Move up the list only when a tested flow needs more from the model - each step costs more per minute, and the rate is shown next to the model in the picker.

​

Backup model

Also under the gear: a backup model, used for the rest of the call if the main model’s provider stops responding.

• It is from a different vendor. What fails is almost never one model - the vendor is unreachable, the key is rejected, the region is down - and every model behind that vendor fails with it. The picker only lists other vendors’ models.
• It latches, like the speech and voice backups. Once the agent switches to the backup, it stays there until the call ends, so a provider that is down costs one pause rather than one on every turn.
• The caller hears a pause, and nothing else. Nothing is torn down, so there is no announcement and no language reset - a language model call is a request and a response, not a live connection.

​

Save and verify

Select Save Changes after changing the model or parameters. Then test a representative path:

• Ask a straightforward question to check latency and tone.
• Give the agent an ambiguous request to check whether it asks for clarification.
• If the agent has tools, test both a successful action and a failed action.

Intelligence affects generation. The agent’s role, rules, and conversation order still come from Agent Core; which languages it may reply in comes from Language; and turn-taking is tuned in Transcriber.