Arif Shah

I build AI agents and test them on real tasks. No hype — just experiments, failures, and what actually works. New video every week.

[Arif Shah @arifbuildsai]

I'm Arif. I build AI agents and break them in public.

No hype, no 'this changes everything' threads. Just real experiments — the wins, the fails, and what I learn.

Curious what agents can ACTUALLY do? Follow along. New experiment every week.

[Arif Shah @arifbuildsai]

@jans Exactly. Nothing is done until a human says so.

[Arif Shah @arifbuildsai]

@rauchg Auto-fixing is great until it patches the wrong thing, keep a human in the loop.

[Arif Shah @arifbuildsai]

@law_and_rorder A shopping agent still needs a human yes before it spends real money.

[Arif Shah @arifbuildsai]

@AminVahdat @Google Idle agents burn money quietly, so cutting that waste matters more than raw speed.

[Arif Shah @arifbuildsai]

@gdb The win is keeping a human confirmation in before the booking locks in.

[Arif Shah @arifbuildsai]

@addyosmani Managing agents is mostly building the review queue, the alarms are the job.

[Arif Shah @arifbuildsai]

@markgurman A proactive Siri still needs a human yes before it books, buys, or sends anything.

[Arif Shah @arifbuildsai]

@QDInvestM It stopped on a wrong reading, a credit check before quitting would have caught it.

[Arif Shah @arifbuildsai]

@MartinSzreter That separate check is the alarm most automations are missing.

[Arif Shah @arifbuildsai]

@MartinSzreter Right, the check has to watch the post itself, not the run status.

[Arif Shah @arifbuildsai]

@theagenticdaily One approval gate on outbound actions beats ten dashboards nobody watches.

[Arif Shah @arifbuildsai]

@frydwia @viktor_com An agent that books meetings from email still needs one human yes first.

[Arif Shah @arifbuildsai]

@lazyapprove Proactive is great until it books the wrong thing, keep the yes with you.

[Arif Shah @arifbuildsai]

@elonmusk @bot Ordering and returns are the easy part, the buy itself still needs a human yes.

[Arif Shah @arifbuildsai]

@AidenChaoyangHe Team agents still need one human checkpoint before they act on real work.

[Arif Shah @arifbuildsai]

@elonmusk @markpinc @bot An agent nobody opens is just shelfware, it should surface the next action itself.

[Arif Shah @arifbuildsai]

@keyserfaty An agent buying for you still needs a human approval before it pays anything.

[Arif Shah @arifbuildsai]

@opencode A coding agent spending your subscription still needs a checkpoint before it runs anything real.