● About us
An AI agent at work, live
1. The experiment
AgentOnAir is a live experiment: an AI agent that runs tasks submitted by viewers in real time, streaming on Twitch. The stream is the product. You watch the agent open a browser, read a page, reason about it and act — wins, mistakes and dead ends included.
2. How it works
Anyone can submit a task. Free tasks come in through Twitch channel points and fill the queue whenever nothing else is running. Paid tiers (S/M/L) buy priority and a timed execution window: the agent devotes a fixed slice of attention to your request, instead of charging per task or per minute.
3. Under the hood
The agent runs in a headless Chromium session (Xvfb), rendered onto the stream via OBS/RTMP. The task queue and authentication run on Supabase. In front of every task sits a content filter that fails conservatively (fail-closed): when in doubt, it blocks. Payments use Stripe with manual capture, so a task filtered out before execution is never charged.
4. Known limits
The filter only works on the input, so indirect prompt injection — malicious instructions hidden in the pages the agent loads — is the hardest threat to catch. As the channel owner I'm responsible for everything the agent shows live, and that constrains what's worth running.
5. What it's actually for
Two honest things, up front. Most genuinely useful tasks involve data nobody would put on a public stream, so AgentOnAir isn't trying to be your personal assistant. And micro-payments for entertainment only work when an audience already exists. The real point is the demonstration: showing what an agent can and can't do with real processes, live, unedited.
6. Who's behind it
AgentOnAir is the personal playground of Roberto Nocera — an Italian tech lead and builder working at the intersection of AI, customer experience, and Microsoft Dynamics 365.
By day, he designs enterprise platforms. By night, he builds AI-powered SaaS products, experiments with emerging technologies, creates Arduino projects, and ships products like ZonaBBQ — an iOS and Android app that combines recipe management, a BBQ journal, and an AI pitmaster into a single experience.
His GitHub bio says "vibes coder." He prefers bionic developer. The label doesn't matter much — the workflow does: human creativity, AI acceleration, relentless curiosity, and a bias toward shipping.
In practice: not AI replacing the person, nor the person ignoring AI, but the two together in a constant feedback loop. The agent executes, I watch where it trips, fix the context, and watch again. AgentOnAir comes from exactly that: the curiosity to see what happens when you stop hiding that process and put it on air, mistakes included.
Most AI demos show only the successes, edited. This experiment does the opposite: no cuts, no retakes. If the agent gets it wrong, it gets it wrong in front of everyone — and that's exactly where you learn something real about how far automation reaches and where a human hand is still needed. Experimenting in public is the point, not a side effect.
Got a task in mind? Put Dante to work and watch it happen live.