Basic Instinct

I've been playing with Instinct lately, an AI agent that can actually operate websites and do things for you, running on whatsapp and I've had a couple of "holy shit" moments already.

Not in the "this changes everything" way, though. I have had enough of those claims for a lifetime, especially when the product in question is hyped up by VCs from Silicon Valley. But I refer to the very practical scenario where I keep finding stupid tasks I have postponed for months, sometimes years, and now I can just throw them at something else while I continue working.

This is fundamentally different from how I use Claude or ChatGPT. I use those to think, debate, research, write code, review documents and generally augment whatever I am doing - but mostly for work-related stuff, so they're well sandboxed. Instinct, on the other hand, gets the stuff I do not want to do myself, with no access to my work stuff, and it's just decluttering my personal life.

This distinction is becoming increasingly important to me.

The first task I gave it was deliberately boring: create a free email account for me on Zoho Mail for my band's website. I wanted to see whether it could navigate the usual signup nonsense without me babysitting it through every click.

It did.

The process lasted for 2 minutes, and it prompted me to click a confirmation email and share with it (instinct) a 6-digit code that was sent to me via email. Needless to say, it doesn't have access to any of my accounts whatsoever.

I've had it buy things from e-commerce websites, too. I asked it to look for discount codes before checking out, because it has infinite time and nearly-infinite resources. I've had it send emails to people on my behalf using its own email account. I've generated a shitload of WhatsApp stickers with it. I've converted documents between different formats. I've asked it to do quite a wide range of tasks and it hasn't deceived me or let me down in any, yet.

Individually, none of these things are impressive enough to deserve a LinkedIn post with a fire emoji and "the future is here" at the top, but the seamless whatsapp ubiquity, the feeling of being able to do almost anything and the familiarity with which it converses on whatsapp (using reactions, quote-replying, etc)... every bit of it is eerily uncomfortable, at times.

The first "holy shit" moment

The most impressive thing that happened was also incredibly mundane: I asked it to buy me a T-shirt just by sharing a screenshot of the t-shirt and my size. It found the product, went through the buying process and, when it reached checkout, all of my details were already there. Name, address and everything else it needed.

My first reaction was: wait, did this thing somehow hack into my accounts? Were my details leaked from somewhere?

Luckily for me, it was neither of the above: it was checking out as a guest and had populated the form with information I had given it in prior tasks - like signing up for the email account or searching for something in my vicinity.

Obviously, there is no magic here. But it still felt like magic because I had forgotten I had given it the information in the first place and because it progressed to the very last step of the process while I was busy living my own fucking life and minding my own fucking business.

That is probably the first time an AI agent has felt remotely similar to delegating something to another person for me. Not because it is intelligent in some philosophical sense, but because I did not have to reconstruct all of the context again. Ask, and it shall deliver. And fuck me if it does not deliver. This thing is pretty wild.

One of the annoying things about most AI tools is that you spend a surprising amount of time explaining yourself or setting them up (Hermes, Openclaw): the same preferences, the same context, the same information about what you are trying to do. Memory has improved this, but it is still mostly useful inside the conversation itself where you actually conduct a big part of your life, which is inside whatsapp. And that's where the magic lies.

Small thing, huge difference.

The best use case is the shit I was never going to do

The task I am enjoying the most is decluttering my personal backlog. One of the best use cases is deleting accounts from stuff i don't use.

I have been online for a very long time, which means I have accumulated an archaeological collection of abandoned forums, SaaS products, websites, random accounts and services I signed up for once between the year 2000 and now, and never touched again.

I occasionally go back to the excel file I have with the logins and delete one or two, but the process is tiresome and sometimes the confirmation emails take hours to arrive, so I abandon the task midway through it.

I was also absolutely never going to complete this titanic task because there are more than 300 accounts left to be deleted, in the first place. Also, there is always something better to do than spend a Saturday recovering passwords and looking for the "delete account" button hidden six screens deep in some forgotten website.

So I gave Instinct this task and that's it. It, very diligently, goes one by one. If it has the password, it logs in. If the password does not work, it tries to recover it. If the website has a normal account deletion flow, it goes through it. If the only way to close the account is to contact support, it sends them an email.

This is exactly the kind of work I want an agent to do. Not my strategy, not an important email to a client, not my blog posts, not deciding what MarsBased should do next year. I want it to do garbage admin so I can focus on the high-level tasks.

This is excellent for the kind of task that is too small to hire someone for, too tedious to do yourself, but still occupies a tiny corner at the back part of your brain because you know it is pending and it has been pending for 16 years. And i've got about 100 tasks like this that require someone or something to tackle them.

The actual superpower is that I can fuck off and do something else

This is where the conversation about AI productivity gets slightly wrong, in my opinion: we tend to measure these tools by asking whether they can do something faster than a human, or even better (in the case of code, for instance).

Can it write the email in 20 seconds instead of five minutes? Can it research something in three minutes instead of half an hour? Can it generate the code of an entire app in a day instead of a week?

While that is a useful and valid way of thinking, there's another angle. The interesting metric is not speed, for me. It is whether I have to be there.

If I can give Instinct a task and then continue working on something else, we are no longer talking about making me 20% faster at the same task. We are talking about parallel work, on top of the feeling of my backlog being decluttered and lifting this heavy burden off my shoulders.

I have limited time, as the CEO of a company, a family man and a fairly sociable person, obviously, but my bigger limitation is attention. I can technically fit a lot of ten-minute tasks into my day if I'm motivated to do them but the problem is that every one of those tasks breaks whatever I was doing before, forces me to change context and leaves some residue behind. Some require to wait for a few minutes for an email or an SMS to arrive.

This is also why I am terrible at doing administrative chores. They are not difficult: they are expensive in interruptions. But not for Instinct. Instinct runs them in the background while I keep doing my work.

I am not outsourcing thinking. I am outsourcing where I place my fucks. And my fucks don't deserve to be placed in admin work.

I have sandboxed the hell out of it

Before you think "look at this reckless bastard!", let me be clear: this does not mean I have given an AI agent the keys to my entire life. Quite the opposite.

For now, I only let it operate on personal stuff. It has no access to MarsBased systems, no access to my company email and no access to my calendar.

MarsBased is currently going through ISO 27001 certification and giving an experimental AI agent access to company infrastructure would be a spectacular way of royally fucking up for no good reason. Huge downside, little upside. Prompt injection is still very real and scares me shitless.

Giving an agent access to email sounds great until you remember that email is basically an inbox where arbitrary people on the internet can send instructions, links and attachments to something that might also have permission to act on your behalf. Same happens with Calendar invites.

No thanks.

Credentials, as of a week ago, are handled through 1Password, so I do not have to paste passwords into prompts or random forms. For payments, it uses Stripe Link, which means it can complete purchases without me handing over my raw credit card details. Instinct doesn't see a thing of passwords/credit cards. It uses established technologies and products as proxies for the secure stuff, and that gives me a relative peace of mind, because so far I'm only buying tshirts and have it download stuff, so it can't fuck my life over in any possible way.

This setup is not bulletproof and I am not pretending it is. What it gives me is a reasonably contained blast radius. If it fucks up buying a t-shirt, I can live with it. If it screws up inside my company email, that is an entirely different category of problem and my partners will be very angry.

The weird thing is that it has not fucked anything up yet

This is where a typical blog post would insert the cautionary anecdote where the agent bought 400 t-shirts, emailed my clients asking for a refund, pinged an ex-gf and accidentally subscribed me to twelve crypto newsletters.

Nope. Nothing like that has happened. No surprises, no fuckups, no scares.

I am assuming that at some point it will make a mistake, and I want that mistake to happen somewhere safe doing something very boring. I assume that this will be more expensive after a while or that its capabilities will be restricted as they scale.

This is probably how I will keep testing its responsibilities: first give it things that are easy to verify and cheap to undo, then gradually let it do more, like onboarding a new hire. This is what I did when I hired a personal assistant a few years ago from an app I had invested in (which was, quite literally, the same as instinct but with a human component).

You do not hire a person and give them production access, your bank account and the password to everything on the first day. You give them a few tasks, see what happens and increase responsibility over time.

Funny how we trust machines more than people.

Where is this headed?

Instinct can send emails using its own email address, which I actually like. I do not necessarily want an agent sending things pretending to be me but there might be cases for that. Also, they're rolling out a Concierge service that'll call for you.

At first I thought it would be cool to make my Instinct bot look more human: give it a normal name, let it send emails like a person and make the interaction more natural, but I'm not all that bought into that.

There is no need to bullshit people. If an email comes from my AI personal assistant, say so. "This is Àlex's personal assistant" is perfectly fine. We will probably get used to this very quickly anyway.

Executives have assistants. Founders have assistants. Soon lots of people will have AI assistants. I don't plan to resist the future. I just want my AI bot to call other bots, where a person is not on the other side (insurance, bank, technical support, etc.). I don't want a bot to call my favourite pizza place two blocks away from me. I want this bot to bug the hell out of that car rental company that ripped me off last summer.

The social contract is the part I'm figuring out now, and it's very interesting and mentally stimulating.

This is where agents make sense to me

I still do not buy the "company of one with 700 agents" fantasy that gets posted on Twitter every day. All of them are bullshit. All of them.

I wrote recently about The golden era of AI-pilled founder entitlement precisely because I think there is a dangerous tendency to mistake infinite execution capacity for judgement. Agents can do more things for you but that does not mean they know which things are worth doing.

I am, however, becoming increasingly much more bullish on a more precise version of the idea: AI as labour for the stupid stuff. An agent that cleans up accounts, fills forms, buys things, follows up on administrative requests, converts files, checks discount codes and eventually sits on hold with customer support for me is already useful enough.

It does not need to be my cofounder or my mentor. It needs to be a roomba for my life.

Every time a stupid task appears now, I am starting to ask myself a different question. It's not "how can AI help me do this faster?".

It's "why the fuck am I doing this at all?".

Disclaimer: I am not an investor in Instinct nor affiliated to it in any way. I'm just a happy user.