Decagon Voice 3 duplex agent or a voiceover job: which do you need?

Pick a duplex voice agent like Decagon Voice 3 when a person talks back to the system in real time; pick a file-based voiceover job when you already know the words. Voice 3, announced October 1, 2026, can process incoming audio while it is speaking, which a script-to-audio job does not do.

What duplex means in Decagon's design

Decagon says Voice 3 removes the delays of a cascaded pipeline, the silent gaps and interruptions that come from passing audio between separate steps. It describes two layers: a low-latency conversational model that listens and speaks, and a more powerful model that handles reasoning, tool calling and guardrails.

That structure exists to survive a live conversation. The caller can interrupt, ask a follow-up, or change topic, and the agent keeps up.

What a voiceover job does instead

A voiceover job has no caller. You send the exact words, pick a voice, and wait for a file. In Sume, the hosted tool is tts_create. It needs an idempotency key, a transcript of 1 to 20,000 characters and a voice selector. The job returns an audio URL, duration, word timings and sentence segments you can feed into a timeline render.

The trade is control for reactivity. Because the script is fixed, you can check it, approve it and reuse unchanged takes. Nothing in the file changes depending on what a listener says.

Side by side

The table below separates the two.

Cost shape

A live agent is priced per conversation or per minute of a service. A job is priced per character. Sume's rate card lists text to speech at $0.0475 per 1,000 characters, with a one-cent minimum per job and a 95-cent ceiling at the 20,000-character limit. Confirm the current rate in GET /v1/catalog.

If you only need a recorded greeting, a menu prompt or a hold message for a phone system, that is a job, not an agent. Generate it once, check it, and load the file.

When to use both

Some teams run a live agent for calls and use jobs for the assets around it: the intro video on the support page, the narrated how-to, the captioned clip in the help center. They share a brand voice only if you pick it deliberately in each tool.