Here are the biggest changes I see coming to AI in 2026.
And when I say “verifiable,” I don’t mean “trustworthy,” which a lot of people confuse and has always been a criticism of AI since 2023.
This is verifiability in the test-driven-development sense. Or the prompting evals sense. It’s testing and judging whether the output is what we asked for.
I tremendously appreciate Andrej Karpathy’s concept of Software 2.0, where he talks about 1.0 being about writing software, and 2.0 being about verifying software.
This rhymes closely with several converging thoughts I’ve had over the last 10 years around goal pursual (TRIOT, 2016), hill-climbing, moving from current state to ideal state, and automation of the scientific method.
Central to all of these is Verifiability.
You can’t hill-climb if you don’t know if you’re getting higher. So the primary questions become:
- What does success look like, and
- What are the verifiable markers indicating that you’re making progress towards it?
I’m obsessed with finding ways to do this across domains. And the cross-domain bit is actually what’s important. One of the main reasons AI impacted coding first is that it’s one of the more verifiable domains that exist.
- You know if something compiles
- You know if something runs
- You know if something produces the output that you wanted
So, what are the equivalents of these for other tasks? Like business tasks and personal tasks. That is what will start to congeal over the next year and after.
Right now it’s not much of a consideration because most people aren’t thinking deeply about AI or how it will be used or how it will apply to their business. They are too busy responding to the board and management telling them to implement it yesterday already.
As great as 2025 was for agents, they were still largely manual and interactive. You ask them to do something, they go off and do it, and they come back with the result.
But if you shut down your interface, like closing down Claude Code for example, it all stops until you start it back up again.
This year, we are going to see a lot more continuous agents via:
- Cloud environments that run in conjunction with your interactive sessions
- Scheduled Agentic tasks
- Triggers for monitoring systems to look for that activate agents to perform tasks based on a particular state change
I’m working on all of these as part of the PAI project. But others will be working on the same things as well.
It will become a lot less valuable to do only parts of a task that you’re asked to do. We’ll all be asked to go vertical, which is a fancy word for solving all the different problems involved in creating the solution.
The new expectation will be that you can go from problem to solution, up to and including the promotion of the solution so it actually gets adopted.
Keep in mind, this is how principals and fellows have always worked. The more senior someone becomes in an org, the more we require them to be vertical, albeit by using their team.
The problem with AI content isn’t only with addictive short-form videos on sites like TikTok, X, and Instagram.
Post and reply-based sites like LinkedIn could become unusable because not only are a lot of the articles going to be AI, but the replies will be as well.
So like, what’s the point? Are we just watching AI talk to AI?
Already, at the end of 2025, when I posted something, I would immediately see 3-4 comments come in within a couple of minutes with these highly articulate, well-formed, and obviously AI sentences and paragraphs. From someone I’ve never interacted with or heard of.
I think this will force us to lower our apertures on who we follow and allow to lock it down only to people that we trust to produce authentic content and actually produce opinions of value.
In 2024 and 2025 it was kind of okay for you to think AI was stupid or hype, and to use it as little as possible at work and in your personal life. The difference between someone using AI and you still wasn’t that large.
That changes this year.
The amount of work that an AI-native person can do will increase so much that hiring managers will be looking at this as one of their primary hiring filters. And even in people’s personal lives, it’s just going to become obvious who is AI magnified and who isn’t.
One of the more positive things I see happening is people getting excited about building and creating things.
In the last month of 2025, much of my friend group has stopped playing games and has started building using Claude Code. And they are addicted.
Addiction isn’t usually a good thing, but when compared to watching TikTok or NETFLIX I have to say this is an improvement. It does present a question, though, of if we’re making things for people to use, who is going to use them?
I think writing a book is largely an orchestration problem.
Plenty of people have amazing ideas for characters or points they want to make across fiction and nonfiction. The issue is being able to hold it all in your mind at once and logically break it into chapters and then churning through the content.
I’m sure there are thousands, and maybe even millions, of people who have lots of notes for a book. Maybe even chapter outlines, or maybe even just a list of ideas for a book they wish they could write.
Platforms like Claude Code and the open-source PAI platform I’m building make something like this a lot more approachable.
My project, and I’m sure many others, will be able to take a whole bunch of notes from you, interview you extensively for one or more hours as you fill in the scaffolding of what you want to happen in the book. If it’s fiction, or what you want to convey, if it’s nonfiction. And then the system will proceed to build out a structure, fill in the main points, and weave the whole thing together in a matter of minutes.
Books are not that long. The issue is just the organization of all those thoughts into something cohesive.
Of course you’ll still need good ideas and creativity, and some measure of discipline. But if the discipline required drops to 2% of what it used to be, and the orchestration component which stopped the vast majority of would-be authors from becoming actual authors goes away, we’re about to have a whole bunch of books hit the market. And some of them will be extraordinary.
If you use X for a particular topic, and especially for AI, it is, without question, the best place to have conversations about the latest things that are happening.
A lot of people who abandoned X to go to Bluesky or Mastodon found out that they were lacking key ingredients and possessing others that were their own type of toxic. Most of the people actually building things and being excited and positive about the world are on X, not there. So the conversations did not have nearly as much energy and positivity to them.
This will largely be an enterprise problem, but it will apply to the internet in general.
So many people are making so many things, and priorities are changing so fast, that they will just kind of stand them up and leave them out there. That’s a whole lot of attack surface that will slowly decay over time.
This will cause problems in two major ways:
- Security is important because there will often be misconfigurations, keys, API tokens, and all sorts of things out there that allow for compromise of accounts and sensitive information.
- People will lose money because they don’t realize how much they’re paying for services that they thought they turned off
Eventually, AI will catch up and agents will be able to track all this stuff down for people. But there’s going to be a window of a few years when this is going to get really nasty. I think that starts spinning up in 2026.