Simon Willison reviews 2026 progress in LLMs and Agents

Yesterday I read an article by Simon Willison: 2026 in LLMs (so far). This is a blog I basically never miss—it's really well written.

Reading this year-in-review, it really rang true..

Simon strings together the important LLM and Agent developments so far in 2026 along a timeline.

Reading it, I felt quite a lot—so many things feel like they happened just recently, but looking back now, more than half a year has already passed. Time really flies.

What I remember most from the start was that at the end of last year, Coding Agents began to grow explosively.

At the beginning of this year, the OpenClaw wave arrived—that surreal renaming journey—which directly made personal Agents take off, and local models were rapidly catching up too.

Later on, there were the discussions sparked by those formidable models like Fable and Mythos, and Agent security issues gradually came to light. The whole industry is really changing fast.

Unexpectedly, a whole year has almost passed in a flash. It really feels like one year now is like three years used to be.

There's one passage in the article that left a particularly deep impression on me.

Simon said that in January he came down with an ailment he calls "AI fever."

Whenever an Agent isn't building something for him, he feels time is being wasted. Even when sleeping at night, he'd be thinking about how he could have had the Agent do a bit more work.

Haha, this description is indeed a bit too real—I know the feeling well. Waiting for tibo to reset—everyone knows what that's like.

Over the past year I've been working on open-source projects and developer tools, and also researching Agent Harness and Skills.

I've had some small achievements—md2wechat-skill passed 3.8k stars, and I've been iterating on a CLI tool for WeChat official account creation for almost a year now.

In recent months, a thought often pops up: I just want to get the Agent running quickly and see how far it can go.

Especially when I see new models making obvious progress again, I can't help but want to try them out—it's just that I clearly feel Tokens aren't enough.

AI can do more and more things, but our time hasn't increased. If every day we're chasing new models and new tools, and in the end we do a pile of experiments but not much truly settles down, that would be a bit of a shame. Models get stronger and stronger, but engineering experience is still important

I've been doing backend development for over ten years. In the past, when developing a project, I had to write a lot of code, look things up, and debug myself—like digging for answers on overflow and trying all sorts of things.

Now with Coding Agents, development speed has indeed increased quite a bit.

Recently I've been planning a teleprompter app for spoken-word recording. At first I thought the features were pretty simple, but when I actually started building it, I realized there were quite a few details that needed to be considered in advance.

If I just let an Agent develop it directly, of course it could quickly produce a version, but some design problems might only surface later on.

So now when I work on a project, I first think clearly about the goals and scope of the first version. When I hit uncertain technical issues, I research them first or do a small experiment to confirm the direction is sound before continuing development.

The skill I use most is obra/superpowers—first for brainstorming, then for making development plans.

Recently I've had a very obvious feeling: quite a few non-programmers in the community have built independent products from 0 to 1, and they often get stuck in a situation where bugs never stop coming—sometimes the requirements direction itself is wrong, and sometimes the execution process goes off track.

So no matter how strong an Agent's capabilities are, how tasks are defined and how execution results are verified still require someone to seriously check things over. To put it plainly, it's leadership and judgment—the intermediate execution process has been compressed by Agents.

I think this engineering experience is still very valuable in the AI Coding era.

Judging whether a requirement is worth doing, deciding how far to take a project first, and how to confirm the results really meet expectations—these things all require us to gradually accumulate experience ourselves. This year, the changes have really been too fast

Simon's article has another big value: it helps us reconnect those scattered technical developments.

Normally, reading the news every day, it's easy to feel that today's model is impressive and tomorrow's Agent has new capabilities. Looking at each event individually, they all seem like small steps forward. But only when you put them on a timeline do you realize how far the whole industry has already come.

I suggest that if you have time, you seriously read it—especially friends who have been using AI Coding and following Agent development.

Many things that felt shocking at the time, looking back now, have already become part of daily work.

There really isn't much need to jump on the bandwagon for every new model and every new tool—protect your attention.

In the last two-plus months of this year, whether it's models or Agent tools, there may still be surprises. At this pace of development, it feels like 2030 AGI isn't impossible either.