Tiberiu Arva

AI architect in banking, building with AI on the side.
🧭 AI tools
🇷🇴 Romanian
🤝 careers

[Tiberiu Arva @artixisonx]

hey, I'm Tiberiu 👋 based in the Netherlands 🇳🇱

day job is cloud & AI architecture at a bank, where I work on making AI agents safe to run in a regulated environment. permissions, approvals, audit trails.

on the side I build small things with AI. a few sites so far, one my wife and I made for learning Romanian. some work, some don't yet.

I post about AI agents, what I'm building, and AI news. happy to learn from anyone running agents for real

[Tiberiu Arva @artixisonx]

source: G2's 2026 AI Agents Report, survey done in July 2026 https://t.co/7XXNMxpaE2

[Tiberiu Arva @artixisonx]

31% of companies say their people spend 6+ hours a week watching over AI agents.

G2 asked 1,019 buyers, and 45% said the checking and redoing eats more than a quarter of the time they hoped to save.

What helps me is starting the agent on read only and letting it change one thing at a time.

[Tiberiu Arva @artixisonx]

Curious where that time goes for people. Is it mostly reading what the agent did, or fixing what it got wrong?

[Tiberiu Arva @artixisonx]

Sunday thought from my 5am sessions.

There's a lot of small stuff I'd never pay someone to help with, but now I just ask AI. A deck before a call, research on a company, a reminder so I don't forget something.

What's one thing you started handing to AI that you'd never have paid a person for?

[Tiberiu Arva @artixisonx]

@sattyyouneed Yes. I ask for a one line fix and get three new files.

[Tiberiu Arva @artixisonx]

@LinusEkenstam @IKEA They will do it eventually on their own. Experimentation in big companies is hard

[Tiberiu Arva @artixisonx]

@alexmichaelio Mine is @bot

[Tiberiu Arva @artixisonx]

@marcusvance_ Oops, did that happen to you?

[Tiberiu Arva @artixisonx]

Claude Code users, what's one thing in your deny list you'd tell everyone to copy?

Reading .env files is the obvious one. I'm after the less obvious ones, the stuff you only blocked after it bit you.

I'll collect the good ones and share them.

[Tiberiu Arva @artixisonx]

@aksenHQ Good one, in addition, I recommend also protect main directly in github to be 100% sure, haha

[Tiberiu Arva @artixisonx]

@aksenHQ Claude used claude key inside env file instead of claude subscription tokens

[Tiberiu Arva @artixisonx]

@TimoBuilds_ New business idea:

[Tiberiu Arva @artixisonx]

@TimoBuilds_ You go me curious, will try it sometime.

[Tiberiu Arva @artixisonx]

@ericson4smith Just submitted mine to App Store review. Used it myself a lot before that.

[Tiberiu Arva @artixisonx]

@joonian Bills are the one I'd be careful with.

[Tiberiu Arva @artixisonx]

@TimoBuilds_ Now I am better at that. It used to happen when I started using my standing desk a few years ago.

But deep focus work, it is much easier for me to do it while sitting.

[Tiberiu Arva @artixisonx]

@TimoBuilds_ I used to find it difficult to focus even when standing at my desk, haven t tried walking yet, hence asking

[Tiberiu Arva @artixisonx]

@diegopetrucci @lawrjones AI or SI? 😂

[Tiberiu Arva @artixisonx]

@TimoBuilds_ How is the focus while walking and working?

[Tiberiu Arva @artixisonx]

sources: Simular's writeup https://t.co/9nHzgum7w1 / the paper https://t.co/xDZJt3XXIp / these are simular's own numbers. same model with its normal tools only got 5.56%

[Tiberiu Arva @artixisonx]

A 27B open model just beat GPT-5.5 at using a computer.

Simular's Agent S4 got 21.3% on OSWorld V2, a test of 108 real desktop tasks. GPT-5.5 got 13.0%. Claude Opus 4.8 got 20.6%

cost per task: $1.10. GPT-5.5 was $25.50

they didnt touch the model. another agent kept rewriting the tools it works through and kept whatever worked

it still fails almost 4 out of 5 tasks, so dont hand it anything important yet