Double Agent

Hi! Dror Globerman here with Skip Intro, on innovation, technology, and you. This week:

1. ChatGPT's agent found me 20,000 shekels in invoices.
2. A venture that once cost millions to set up, now costs a fraction. What's yours?
3. We generated the music video while playing in a band rehearsal.
4. Who is this AI agent really working for?
5. The world fears AI's dangers. That's Israel's opportunity.

I'm writing these words with one eye on the other screen.

Because on my left-hand screen, ChatGPT is at work right now, this very minute, saving me thousands of shekels. All on its own, it's going through every credit-card charge since January, visiting the sites I bought from, downloading the invoices, uploading them to my invoicing platform, Morning, and neatly organising the lot for my accountant. It has already spared me at least five hours of Sisyphean drudgery, or 20,000 shekels' worth of lost invoices.

At exactly the same time, Claude is reading the research notes and line-up for my Radio show, cross-referencing them and drafting a suggested script for each interview. That's after it had already put together my weekly news round-up from sources I'd picked in advance.

The talk I gave in Bucharest last week? ChatGPT has already edited the video for me, adding subtitles and fixes.
Then I remembered I had footage of other talks I'd never got round to editing and uploading.

Until a couple of months ago, that was a question of time and money. Not any more.

"Take my talk from the IAC conference in Miami last January," I told it. "I can't remember where the files are. Find them somewhere in my Drive. Stitch them into a single video with subtitles and fixes. Find other angles of the talk online and weave them in nicely. Add subtitles in English and Hebrew."

It took 20 minutes, during which it also tracked down extra footage and edited everything beautifully: cutting to the slides when needed, then to the audience applauding. Then it uploaded it to YouTube. Add all the details underneath, I asked. Done. "Actually," I thought, "tidy up every video I've ever uploaded to YouTube." Sorted.

That's what my days have looked like for the past fortnight. A cascade of fixes and aggressive streamlining of how I work, sometimes two or three improvements a day.
I don't go to bed without ChatGPT and Claude working on something overnight.

I didn't sort out a few settings in my newsletter service, Mailchimp. I asked Claude. In it went and sorted them.

ChatGPT's new agent, dots, now routinely combs through the whole family's emails and calendars and flags anything urgent. Yesterday it warned that my daughter Dani has just a quarter of an hour to get from school to her singing lesson. Worth checking she makes it. Genuinely, thank you.

But that's only half the picture. Replacing me in things I used to do (somehow) is nice, but hardly groundbreaking.

There's another half: tasks I don't do and never dared imagine, because I lacked the ability, the time, the talent or the money.

I've had an idea for a venture for ages. With AI tools, I've mapped out the business process and built a working prototype website. That's the trivial part. 

Because now I can devise a strategy that simply wasn't economical until very recently: building a separate app for each of the hundreds of different uses my idea has. It fits how I imagine people will use it, but not long ago it would have cost me millions of shekels. Now it's a few hundred shekels, perhaps a few thousand if I really go to town.

Welcome to the era of agents that actually work.

"Two armies line up for battle, and Yuval moves between them, sharpening knives for both sides."

We fell about laughing.

Hang on, let me explain. I play in a rock band led by singer-songwriter Yuval Gurevich. When he isn't making music, Yuval sharpens knives for a living. Really.

We're about to release a new single to radio. Last night, in the middle of rehearsals, it suddenly struck me: why don't we make a video for the song?

Anyone got an idea?

Our bassist, Tal Markus, came up with a stroke of genius: "Armies lining up for battle, and Yuval moving between them, sharpening knives for both sides."

I described it to ChatGPT, asked it to write a script, and gave it a photo and a video of Yuval. Off it went. I was playing keyboards, with my phone propped on top of them making the video at the same time. A few more tweaks, and that was that. Perfect? No. But a moment ago it wasn't even an option. For the first time, I can actually ask the band: anyone got an idea for a video?

So now everyone has agents doing things for them.

The great promise of AI is to save us time. In my experience, something slightly different is happening so far: I work roughly the same hours, but I cram far more into them. Videos that sat waiting months for an edit suddenly go up on YouTube. Things I'd have put off until I found someone to help simply happen. Wonderful, right? Except my competitors are getting the same superpowers. All this productivity may not buy us more free time at all; it may simply become the new bar we're expected to clear.

I'm still asking AI for too little. I asked it to edit a video, then to write a description and tags for YouTube. Then I raised the stakes: do it inside YouTube. And again: do it for all my previous videos. We're so used to being the computer's hands that it takes a while to accept it could become ours. The gap between "I want this to happen" and "it happened" is starting to close.

But the moments when the agent genuinely surprised me came without my asking for anything. The speaking invitation I hadn't answered, the graphics change I'd requested that never happened, the clash between my daughter's after-school activities. Things I didn't even know I was missing.

Having felt the benefit, I find it hard to imagine switching it off. That's probably how dependence on technology takes hold.

Who is this agent actually working for? If I'm paying it to look after my affairs, I understand the deal. When an advertising company offers to run my life, I want to know how it makes money from that. Access to my email and calendar, let alone to devices that see and hear the world around me, gives it far deeper reach than it ever had through the feed. For me, the question of who runs the agent, and on what business model, matters at least as much as how clever it is.

Who do I belong to? If my agent decides where to order flowers, which hotel to book and which software to use - who owns the relationship with me as a customer? I caught myself working on an idea for a service, agonising over the copy for the website, and suddenly stopping: who is actually going to read it? The most important visitor to the site may well be an agent that needs to understand what I'm offering, what it costs and why it suits its user. Businesses will have to learn to explain their value to readers who aren't impressed by a pretty homepage.

More on agents in the latest episode of "While You Were Working" (in Hebrew), which I recorded with Danny Peled.

Last week I grumbled here that we non-Americans don't get a seat at the table where Trump and the AI industry's top brass are deciding the world's fate. We Israelis are exposed to all the dangers (models slipping out of control, hacks both malicious and accidental), which grow by the minute, yet we have no say in the policies that count. How frightening!

Yet this very intersection of AI and cyber risk hands Israel a fantastic opportunity. In AI we may be lagging a little, but in cyber? Oh ho, that's our home turf! Bring on the threats, the breaches, the wars.

"How do we carry Israel's comparative advantage in cyber over to AI?" I've chaired perhaps 13 panels with that title at conferences of every kind. And now the world has gone and done the carrying for us.

This isn't really new. Israel is already a leading global hub for AI security: not merely using AI to detect attacks, but protecting AI itself. Israeli companies are building mechanisms that prevent data leaks, spot malicious instructions hidden in documents and websites, and rein in AI agents that might act beyond their permissions or against company policy. Precisely the things the world is afraid of.

There's already serious money here. SentinelOne announced its acquisition of Israel's Prompt Security in a deal reported at around $250 million; Noma raised $100 million in 2025, and Zenity raised $125 million in August 2026.

So how prominent is Israel, and where does it stand globally? According to the industry map maintained by Prompt Security, Israel ranks second only to the US, with 104 of the 451 companies in AI and agent security, roughly 23%. That isn't an official market share, but the picture is clear: Israel is already one of the leading centres for ventures that protect us from AI's harms.

And this is only the beginning. Soon AI will be unusable without protective tools like these. Economically, perhaps even legally, it won't be possible to run it without passing cyber controls. Israel's opportunity has never shone brighter.