AI Agents

AI agents (also called flows) are the reusable instructions behind every HUM call. A good agent gives the AI clear goals, boundaries, fallback behavior, and follow-up structure.

Creating an agent

Go to AI Agents → New agent. The core fields:

• Name — required label shown throughout HUM.
• Kind — inbound (answers calls) or outbound (places calls with task context).
• Greeting — the opening phrase the caller hears.
• Instructions (system prompt) — the agent's primary direction: who it is, what to collect, what it may promise, and when to hand off to a human.
• Voice — pick the voice the agent speaks with.
• Voice model — choose a model for this agent without changing your other agents.
• Voicemail — whether the agent can capture a voicemail when a transfer fails or after hours.
• Recording announcement — whether the agent announces that the call is recorded.

Writing good instructions

• State the agent's role and company in the first line.
• List the specific facts to collect (name, callback number, reason, urgency).
• Spell out what the agent must NOT do (quote prices, give medical advice, promise callbacks it can't schedule).
• Describe when to transfer and to whom ("billing questions go to the billing line").
• Keep it under a few hundred words — shorter prompts are easier for the AI to follow consistently.

GPT-Live-1 calls

Select gpt-live-1 (full duplex) under Voice model for an inbound or outbound voice agent. Existing agents keep their current model until you change it. New agents continue to default to GPT-Realtime-2.1.

GPT-Live-1 listens while speaking and handles interruptions naturally. It adds new voices, including Quartz, Willow, Gleam, and Meridian. Voice choices follow the selected model; switching back to Realtime resets a Live-only voice to Marin.

Use the existing System prompt or Task prompt for the assistant's role, language, tone, business rules, confirmation requirements, and tool workflows. This one prompt guides both the speaking assistant and its reasoning backend; there is no separate conversation prompt to maintain.

Reasoning model chooses the backend that handles those business rules and connected tools. GPT-5.6 Terra is the default; GPT-5.6 Luna is a lower-cost alternative, and GPT-6 Astra is an optional higher-cost model. GPT-Live-1 still handles speech regardless of the reasoning model. Leaving this setting at the organization default inherits your organization's configuration. Voice and reasoning usage are tracked separately in the call's cost estimate.

Test the agent before moving phone traffic to it. Model-generated greetings and recording notices are not guaranteed to be word-for-word; workflows requiring exact, uninterrupted disclosures need verified audio playback. MCP integrations need a Streamable HTTP server URL for Live; connector-only integrations must remain on Realtime. Existing approval policies still apply.

Transfers

Each agent has a transfer map: friendly labels (like "front desk" or "billing") mapped to phone numbers. The agent can ring those targets mid-call, and falls back to voicemail when nobody answers. Agents can also transfer a caller to another AI agent — useful for a main-line agent that hands callers to a specialized sales or support persona without dropping the call.

Business hours

Set a weekly schedule and timezone per agent. After hours, the agent stops offering live transfers and routes callers to voicemail with an after-hours message instead. This is enforced server-side, not just in the prompt.

Tools

Agents can use tools during calls: knowledge lookup, transfers, voicemail, touch-tone (DTMF) navigation, sending an SMS to the caller, and any HTTP tools your organization has added from the Tool Registry. Bind tools per-agent so each agent only has what it needs.

Testing changes safely

Use the built-in test call feature to talk to your agent from the browser before pointing real traffic at it. For bigger changes, duplicate the agent, edit the copy, test it, then re-point the phone number once you're happy.

SMS and event agents

The same agent concept drives SMS conversations (set the agent's channel to SMS — see SMS AI Agents) and event-triggered automations that run when something happens in HUM (e.g. summarize and notify staff after every completed call). Event agents can also be fired from your own systems over HTTP, with the result delivered back via a signed callback — see Trigger AI Agents via API.