Hey Product Hunt,
We're Khaled and the Neural team. Today we're launching NeuralAgent 2.5, the biggest update we've shipped.
For those who don't know NeuralAgent: it's an AI agent that sees your screen and controls your computer. It opens apps, clicks buttons, manages files, and handles tasks across your entire system. You describe what you want in plain English and watch it work.
Here's what's new in 2.5:
NeuralAgent
Hey Product Hunt 👋 Khaled here, founder of NeuralAgent.
NeuralAgent is an AI that actually uses your computer, it sees your screen and controls the mouse and keyboard to get real work done across desktop and browser apps.
What's new in 3.0: this is the release where NeuralAgent gets fast. Our previous versions were about being able to use your computer, 3.0 is about doing it at lightning speed, powered by our new purpose-built Fast model.
We built the Fast model for one job: executing UI actions, clicks, typing, navigation, at ~285ms each. It's paired with a reasoning model that plans and supervises, so the Fast model flies through routine work while staying smart enough to recover when something goes wrong.
⚡ Quick update: it's gotten even faster since the launch video above was recorded! Here's a more recent one, NeuralAgent sending a WhatsApp message in real time. Keep an eye on the cursor, that's the Fast model executing at full speed. In the normal flow, it plans the task once and lets the Fast model fly through the execution, only re-planning when something goes wrong and the supervisor steps in to recover.
Computer-use agents have been useful but slow, you sit there watching them think, click, and wait. We wanted it to feel instant.
Also new in 3.0: Fast Model Replays, do a task once, save it, and re-run the whole thing in seconds by text, voice, or a chat mention. Important part: a Replay is not a recorded macro. It doesn't fire saved coordinates like RPA, on every run, the Fast model re-analyzes the live screen and grounds each action fresh, so it adapts when the interface changes instead of breaking like a brittle script. You get the lightning-speed of a saved workflow with the resilience of a model that actually sees what it's doing and a supervisor that makes sure that the task was executed. And Smart Model Routing sends each step to the right model, so even long multi-step tasks stay fast.
And this sits on top of everything NeuralAgent already does, Watch & Learn to turn the way you work into reusable workflows, a library of Skills (Google, Excel, PowerPoint, PDF, coding, research), scheduled workflows, and Enterprise cloud computers for running UI work at scale.
A bit of context: NeuralAgent is used by tens of thousands of people around the world, and we're genuinely excited to launch this today.
The demo above is real-time. I'd love your feedback, and I'll be here all day answering everything 🚀
For more Fast model demos:
https://www.youtube.com/channel/UCXGEZyKZZMZzOkTAA_r5TVA
The replay architecture is the most interesting thing here. Not-a-macro is doing a lot of work in that claim though. If the Fast model re-analyses the live screen on every replay run, what's the latency cost versus a recorded macro when the interface hasn't changed? And when it does adapt to an interface change, how does it signal to the user that it deviated from the original task path rather than silently doing something adjacent?
The brittle script problem is real. Curious how you surface the cases where "adapted" actually meant "got it wrong."
Love the Fast model plus reasoning model architecture, I am a little curious about, when the supervisor steps in to re-plan after an error, how much latency does that recovery typically add to the flow?
A lot of computer-use demos still feel like watching a slow remote desktop session. The Fast model + supervisor split is the bit that makes NeuralAgent feel practical, not just impressive in a demo.