Engineering
What we fixed after pitching ORION live
· 4 min read · Sheikh Mohammed Saqib
The pitch
From 2 to 4 November I was in Sharjah for the UAE Hackathon 2026, and I pitched ORION live. Not slides about what ORION might do one day — real desktop tasks, on a real Windows machine, in front of people.
A live pitch is an honest test. If the agent gets something wrong, everyone sees it. That pressure is useful, because it shows you exactly which parts of your product you trust and which parts you only hope work.
The bug that mattered
During prep, I asked ORION to create a meeting agenda document and save it to my Desktop. ORION replied: verified, saved to your Desktop. It sounded perfect. It was wrong twice.
The file was in a different folder. And when I opened it, it didn't contain an agenda at all — it contained instructions about how to write an agenda.
Here is why. The document tool only checked that a file existed after writing it. That's not verification; that's checking that something happened. And the sentence telling me where the file was saved didn't come from the tool at all. The chat model wrote it, because that's what a helpful reply usually sounds like.
So the fix had three parts. First, ORION now saves where you ask, using the real Windows folder location instead of a guessed path — so Desktop means your actual Desktop, wherever Windows keeps it. Second, the document tool rejects drafts that are instructions instead of content, or that drift off the topic you asked for. Third, and most important, the reply you see is now built from the tool's evidence — the real path and what was actually written — not from the model's words.
Desktop control v2
After the hackathon I rebuilt how ORION controls the desktop. It now types in any language, including Arabic and emoji. It pastes long text and then restores whatever was on your clipboard before. It clicks buttons by their name rather than by guessing screen positions.
It also knows when to stay out. ORION refuses admin and protected windows it can't safely control. It stops if you switch windows in the middle of a task, so it never types into the wrong place. It never fills passwords, PINs or one-time codes. And it asks you before pressing pay, delete or send-style buttons.
The principle
All of this comes back to one rule: ORION never says done until it has checked.
An agent that reports success it hasn't verified is worse than one that fails loudly, because you stop looking. I would rather ORION tell you "I couldn't confirm that" than give you a confident sentence that isn't true.
What's next
Next, I'm working on element-by-name planning for small local models, so ORION can plan clicks by what a button is called even on modest hardware. I'm also making input faster, and testing ORION on more real tasks — the everyday kind people actually need done.
ORION is in early access and still growing. If you want to try it, request early access on the contact page.
