📩 texting @agent is all the rage

DRAFT · 21 September 2026 · Edition number to confirm. Shaded notes are unfinished editorial items.

Yo yo folks,

TLDR for this week:

• A few more deals closed, and education is paramount.

• Faster AI, more personal agents, and the argument for slowing everything down.

• A way to take your skills between Claude, Codex and other agents.

• And thoughts on Apple’s folding phone, and the new paradigm

1. What’s new in the business

We closed a few more deals this week, which makes the Bali trip rather productive.

Something I’m noticing is an increasing number of people who just want education, or require education as part of the implementation.

Which has proven to be an interesting distinction for us.

There’s the work of implementing something. And then there’s helping someone to implement it, which is often just as time-consuming as the build, but arguably more valuable as you’re giving them autonomy to build themselves in the future.

A thought from Bali

I have so many thoughts on Bali. The most notable of which is that it seems to implicitly cater to two different versions of society.

If you come from money, nay, if you come from an OECD nation, or a nation of decent wealth your buying power is just so much greater than the local economy.

This lends itself to a couple of distinctions:

1/ it attracts a certain crowd (I’m wary of the irony)

2/ it empowers people who would otherwise not feel so high and mighty to act all high and mighty (in not the greatest way possible).

I will be writing more on this at the conclusion of the trip so stay tuned.

I will say, however, the food has been exquisite (touch wood here I don’t get Bali belly with 3 days to go), the people are lovely and the resorts/wellness centres/villas are stunning.

🤖 2. What’s new in AI

The creator of ChatGPT just launched his latest startup

Jev is this new AI tool created by one of the creators of ChatGPT.

Twitter tweet

Think of it as a more deterministic LLM.

When ChatGPT answers something for you, it uses tokens in the background to give you an answer back.

Jev works to just give you robust tasks much quicker.

Its speed is about 20-200x quicker than ChatGPT, although I think it’s largely unfair to compare it against an LLM, as it aims to solve different problems.

Ultimately, Jev is built to make structured decisions inside software. You supply information and define the questions or possible outputs; it returns decisions with probabilities and confidence estimates. TypeSafe (the company behind it) calls this a “System One Model”.

Imagine processing a customer email. You could ask:

• Is this about billing, cancellation or technical support?

• Does it need urgent attention?

• Should a person review it?

Those are the kinds of decisions Jev targets. Writing a thoughtful reply to that customer is a different task and one that an LLM tackles. You can use it here

And see a really cool use case here (it’s so fast)! (this is the best example of how fast it is)

So (when) should you use it?

You’d use ChatGPT for an open-ended conversation: ask questions, explore ideas, write something, then refine it.

With Jev, you define a specific decision you want made:

• ChatGPT: “Read this customer enquiry and help me work out how to respond.”

• Jev: “Classify this enquiry as education, implementation or both.”

Its playground lets you experiment manually, but the main use is embedding those decisions into software, especially when you need to assess thousands of enquiries using the same criteria.

This would be really helpful for big enterprises that use a lot of tokens up on mundane tasks.

But for the everyday person, needn’t worry about it.

Personal agents keep kicking

three core agents to cover

Instinct is the personal agent that runs your life proactively.

Instinct is a personal AI agent you just text on your WhatsApp or your iMessage. You connect it to your apps and then it learns from you and runs proactively over time.

Hilariously (at least for me) it’s basically the equivalent of what the vision for AxleClaw was (which we’ve largely reserved for internal use as we realised to make it really really good would require venture money).

It’s invite-only which has proven itself to be a rather scalable acquisition feature apparently.

Getting an invite makes it exclusive. They deliberately targeted more exclusive high net-worth individuals and people in the know which in turn created buzz around the product.

I got sent it the other day but to be honest I was a little let-down (at least for now).

I’ve seen some cool use cases for other people but it hasn’t been all that proactive (as of yet).

I feel like a lot of these tools provide a cool “oh wow” moment for a lot of people who haven’t seen it before but after 30-days of using it - I think it would be hard for most people to come to terms with paying for it.

Unless it’s a business use-case like Grok Bot (which has been on absolute tear in terms of their marketing - looking to do a deep dive on that next week).

But as Paul Graham would say for startups, make something that people want and then worry about monetising it later.

For most consumer products ads is how you would have to monetise, but because it’s inside of iMessage and WhatsApp it may be tough to actually do so.

It also doesn’t solve a specific pain point, per se.

But hey, I’m not raising at a $10b val so what do I know?

Meta launches Muse

Meta has also launched Muse, a personal agent that people can contact through its own app or WhatsApp. Initial availability is in the US. Meta’s announcement

This one is superrr interesting because Meta basically knows everything about you (if you use the socials at least).

They don’t necessarily need to make money off of this (in the short run at least) - they just want to collect more data from you to feed it back into the ad machine.

But there does seem to be a bigger play for them.

They could be looking to become the app store for agents.

They already have massive distribution. Now they want developers to come and flood their connection to provide plugins and tools for the agents to call upon so that Muse can use them to fulfil on the actions that users want.

Twitter tweet

Another super interesting part of this is that they seem to be going after a different market.

See below for how Meta is going after TV commercials.

The older generation who already have WhatsApp, Instagram and Facebook installed and know how to use it, but don’t know how to use AI or are scared to use it, or don’t even know it exist (yes there are people who don’t know AI exists).

Twitter tweet

See the post from Alexandr Wang (founder of Scale who was acquired by Meta) above.

Going after this older demographic will provide a really interesting beachhead.

The AI Agent you wear on your wrist

I’ve personally been thinking about what the next paradigm of technology looks like.

It’s anything that removes us from screens. It’s really unnatural that I just sit here typing things. Or I’m constantly pulling out my phone to type something mundane out. (more on this below)

Introducing Persona.

It’s an AI agent you wear on your wrist which you can speak into. It’s plugged into your phone and like Instinct proactively does stuff (e.g. if you get off of a flight it should know automatically to book you an uber home). The idea is that you can just speak to your wrist and things just happen.

Twitter tweet

For reference this is created by the founder of Cal AI, Zach Yadegari (who is just 19).

I think this is so cool and he can probably pull it off, running off of the back of the momentum from selling CalAI for $250mil to MyFitnessPal.

Also young enough and emboldened enough to see it come to life.

And I think I would actually use something like this - I’ve been saying this to Swanny for ages that I think we’ll look back in 100-200 years (provided we’re alive) and question why the hell we spend 8-10 hours a day sitting at our desks doing mundane tasks.

Bullish on this, keen to see where it goes.

Everyone wants to slow the frontier

Dario Amodei published an essay arguing that the development of more capable AI should be paced so that safety work has time to catch up. He also discusses the difficulty of doing that while maintaining a lead over China. The essay

This all comes after the Jacob Coxon tweet

And is one of the rare times the frontier labs have all agreed on something.

It’s worth engaging with the actual argument.

If the consequences of a mistake become more serious as systems become more capable, there’s a reasonable case for demanding better evidence before deploying them.

Then come the difficult questions.

Who decides the pace? What evidence changes that decision? Can smaller companies meet the requirements? And how do you stop rules written around today’s leading labs from protecting their position?

Having a commercial interest doesn’t make someone wrong. It does mean their proposal should be examined beyond the stated intention.

I don’t have much more to add after last week but it’s something worth thinking about again.

I spoke about this in depth last week

A few clips to share

There were some great clips I found from Twitter this week (one from the All-In Summit, including this one from one of the hosts going at Meta and one on Neuralink - which btw is fkn insane and so cool).

Twitter tweet

You have to click on the above to make actually see the clip but trust me it is so cool.

Twitter tweet

It’s rather refreshing to see people hold big tech accountable, kudos to the host Jason here.

👷 3. Builder’s tools

Managing Claude usage

One of the biggest things I hear is "when should I use each model?” and “how do I manage tokens as an organisation on claude?”

Claude released a video which outlines just this.

The idea is that you should use Opus or Sonnet for more mundane tasks and then something like Fable for higher level thinking (see video for more detail).

If you run an organisation, you can actually toggle certain models on and off for specific model usage. E.g. if you only want some people using Fable you can set it to be this way.

Your skills can travel with you

Notion’s new Skills API is a useful development.

Teams can maintain skills in Notion and export them in standard formats for use across other agents. Notion describes integrations for syncing skills to GitHub and installing them through Vercel’s skills CLI. Notion’s announcement

A skills library for every agent

That means the process you’ve worked out for writing a proposal or reviewing a document can live somewhere your team can edit it, then be used across tools such as Claude Code and Codex.

There’s a distinction here: you’re moving the instructions. You aren’t automatically moving every conversation, memory or connected account.

Still, this is very useful.

If you’ve spent time teaching an agent how your business does something, those instructions are an asset. Keeping them accessible and portable gives you more freedom to change tools without rebuilding the process from memory.

And Notion has enabled this - all without you having to actually worry about version control and what github is.

Can you reverse engineer taste?

Taste Labs is another company I found interesting.

It describes work on making design preferences and subjective judgement usable by models and agents. Its Brand API includes extracting brand systems, finding design inspiration and checking brand consistency.

“Reverse engineering taste” is what they’re going after.

You can give an AI a technically correct brief and still get something that feels completely wrong. The colours might match. The logo might be there. The result can still look like nobody made a deliberate decision.

I’m curious how much of that can be captured through examples, comparisons and feedback. I got recommended this by a friend only today so will report back next week to let yall know what I think.

🤔 4. Differentiated take: what comes after the screen?

Apple announced the iPhone Duo, its new folding phone, and I’m not sure how I feel about it.

It seems super cool. But what’s the market for it?

Is it a smaller iPad? Is it a bigger phone?

I do love the ambition and the fact that they’re trying something new.

There’s an argument for having both in your pocket. More room for reading, watching something or working across two apps, which folds away when you’re finished.

But is much harder to justify at a $1999 price tag.

Availability is scheduled to begin on 23 October, including Australia. Apple’s announcement

See the new Apple CEO shaking introducing it to iShowSpeed (buddy was nervous).

I can see why someone would want it.

But one of the difficult things about trying to innovate on the phone is that the medium is already so good.

A rectangle that fits in your pocket, with a screen you can touch, turns out to be an incredibly effective way to do a lot of things.

Making something meaningfully better is hard.

We’re part way there.

E.g. I’ve been plugging my laptop into charge and walking around in the sun, talking to ChatGPT or Claude about what I’m working on.

It’s still kinda clunky though. Plus I still need to review everything at a screen.

Having said that, I don’t necessarily need a screen for every step that gets me there.

That’s kinda what Siri was meant to offer.

Say what you need, and have the device help you do it.

The difficulty is everything behind this. I.e. it understanding you. Having the right context. Knowing when to act and when to ask. Getting it right consistently enough that you stop checking every step. And it having access to enough tools such that it can actually do the work that you need.

Until that works, tapping through an app can be easier than explaining yourself three times.

Glasses introduce another possibility: an assistant that can see what you’re referring to, instead of requiring you to describe it first. There are some cool companies doing this - Meta is doing this too but I don’t think most people actually trust Meta enough (we’ll see though)

Neural interfaces take the idea much further, although that feels like a longer-term question (e.g. Neuralink) and not sure how the everyday person feels about having a chip in their brain.

I don’t think screens disappear, maybe they become more virtual? And more futuristic like Iron Man or something from Tomorrowland?

After all, often looking at something is simply the best way to understand it.

But I suspect the next substantial change will come from needing to operate our devices less.

How much of what we currently do on a screen should require us to open it in the first place?

And what does the next 20 years look like?

5. Carve-outs

• the acquired podcast stays undefeated: who knew Home Depot was the greatest stock return of all time?? (bigger than Apple’s) - go listen if you haven’t

• music: tns (is an absolute track) and innerbloom stays goated

• plus I recently downloaded X and would love to connect if anyone else is on there! https://x.com/alexbsidhu

Adios amigos,

Alex