GPT-Live gives Picsart AI Agents a real voice: you talk, the agent talks back, and the creative work keeps running in the background, even after you hang up. That combination is the story OpenAI chose to spotlight in its GPT-Live startup showcase, where Dmitry Ivanov, Senior Staff Engineer at Picsart, walks through how the team built it. Voice arrived in Picsart AI Agents in September. This is the behind-the-scenes look at why it feels so natural to use.

A brand born from one conversation

Picture this. You have had a half-baked idea sitting in your notes for weeks: a coffee subscription called Slow Lane. You open Aura and simply say, “Remember that coffee idea we talked about?” Aura does remember, pulls up the earlier work, and suggests starting with the look and feel.

Then the conversation does what real conversations do. You want the tone to feel unhurried. You suddenly wonder out loud whether the idea is any good at all. Aura answers the doubt honestly, then offers a small, low-pressure first step: a visual direction. Moments later a board appears with colors, a serif name, a tagline, mood shots, and a tiny playful snail.

You ask for an Instagram launch kit. Aura keeps you posted by voice as the pieces land, and the call ends with three feed posts, two stories, a vertical reel, and captions and hashtags saved and ready. Not a single prompt typed. That is the demo Picsart shared with OpenAI, and it runs on Aura, the creative director agent that brings in the right specialists for each job.

Dmitry Ivanov on why voice belongs in creative work

In the showcase, Dmitry Ivanov, Senior Staff Engineer at Picsart, explains where the idea started, with a simple observation about how ideas actually happen:

“Creative work can outlast a call. We wanted users to talk naturally to Picsart AI agents while a task runs, or leave the call and return when the results are ready. Creative ideas don’t always arrive fully formed: talking through a mood or rough direction can be easier than writing a precise request. GPT-Live was the first technology that unlocked this vision for us. Its natural live conversation made voice feel like a way to develop ideas together, rather than simply dictate commands.”

Here is what each of those ideas means once you start talking:

  • Creative work can outlast a call. A launch kit takes longer to make than a quick chat lasts, so the work never depends on you staying on the line.
  • Talk while the task runs. Keep shaping the brief out loud while the agent generates, instead of staring at a loading bar.
  • Leave and come back. Hang up, grab lunch, and find the finished work waiting in the same thread.
  • Fuzzy ideas are welcome. “Something calm, a little playful” is a perfectly good starting point. The agent asks the follow-up questions for you.
  • Create together, not command. GPT-Live listens and speaks at the same time, so it feels like brainstorming with a collaborator, not reading out a to-do list.

The secret: talking and making are two different jobs

Most voice tools try to do everything inside the conversation, so when the talking stops, the work stops too. Picsart splits the two. GPT-Live runs the live conversation, while Picsart’s own intelligence layer does the heavy lifting behind it: reasoning, picking tools, planning, and remembering your saved context.

The audio streams directly between your browser and OpenAI, which keeps replies quick and natural. Meanwhile, the Picsart app passes your requests to the agent and carries progress and results back. Each call also starts with the instructions of the agent you picked, so a chat with Mira, the brand designer, sounds like Mira.

Voice, typing, and on-screen clicks all share one thread, one agent, and one set of approval rules. Start a project by talking on the go, finish it by typing at your desk, and nothing gets lost in between.

Why the agent never talks your ear off

Ever used a voice assistant that narrates every tiny step? Picsart designed against that. The team’s rule is that voice should know more than it says, so the agent receives three kinds of information, each with its own job:

  • Instructions: who the agent is, when to hand work off, and how to hold a good conversation.
  • Thinking: quiet background context, like the latest chat, how each task is going, and what you just tapped on screen.
  • Commentary: the answers, questions, and big updates worth saying out loud.

The result is a calm, focused conversation. Routine steps happen silently, while “here’s the plan” and “your reel is ready” get spoken. Replies arrive as full sentences at natural pauses, so the agent never trips over its own words. Your requests are protected too: each one is saved with its context the moment you send it, so a stray comment later can’t quietly rewrite a job already in the queue.

Hang up anytime, the work keeps going

Interrupt the agent mid-sentence or end the call, and your task keeps running. Stopping a job is always a deliberate choice: you ask the agent to stop it. New requests wait their turn instead of scrambling work that is already in progress, and the agent only says work has started once it really has.

Spending stays in your hands, too. The plan, its cost, and the question asking for your go-ahead always come as one package, and a passing comment never counts as permission to spend more.

And when you come back, the agent picks up right where you left off. Each new call loads your latest messages, attached files, running jobs, and anything still waiting on your approval. Ask it to “use the style we saved,” and the agent looks it up for you.

Building voice agents with GPT-Live: 4 lessons from the build

Building something similar? Dmitry closes his write-up with four lessons, and they work as a checklist for any voice-first product:

  • Prepare answers for voice in the main agent. Short, conversational replies win. Keep it to a single question per turn, and spell out what will happen, what it costs, and what needs a yes before the voice speaks.
  • Save each request’s context when it is sent. Lock in the context the moment a request goes out, so later chatter can’t change a job mid-flight.
  • Share the thread and rules across voice, text, and screen actions. People switch modes constantly, so every mode needs the same agent and the same approval rules.
  • Keep the conversation and the task separate. Calls end and tasks finish on their own schedule. Neither should ever break the other.

Get answers to common questions

Picsart is one of four startups in OpenAI’s “Pushing voice agents further” showcase, published November 10, 2026. Dmitry Ivanov, Senior Staff Engineer at Picsart, explains how Picsart keeps voice conversations and agent tasks in sync, with a demo of a voice session with Aura.

Say it out loud and watch it come to life

That idea you have been meaning to write down? Skip the writing. Open Picsart AI Agents, pick a specialist or let Aura lead, and start talking.