Adam Keys is Thinking

Everything is weirder and more capable. Models are way better at many things, unattended coding in particular. But, to everyone’s dismay, that causes them to write in a tedious style of English.

There’s plenty of room for humanity in between the tokens.

1. Models & harnesses

Frontier lab models, despite their quirks, feel like they will graduate to boring technology soon. When the servers are healthy, you can rely on Anthropic and OpenAI’s models to get work done. Cursor’s in-house models (probably other labs too, I haven’t tried Kimi, GLM, etc.) are solid enough to pinch-hit for frontier models when the inference servers are buggy.

Harnesses, when they’re relatively bug-free, are basically just as reliable. You can use the frontier labs (I’m still on Claude Code, mostly) to get most sorts of stuff done. You could substitute Cursor, Pi, or most of the others to personal taste. And, personal taste is increasingly a thing across the harnesses. Claude Code is very much for the terminal people, Cursor is increasingly for the holistic product people, the others are for the customization enthusiasts.

Occasionally one model will make a noticeable leap over the others in capability, speed, or “workhorse-ness”. Occasionally, a model gains a unique capability that is a big advantage until it’s cloned into the others. I find those momentary blips don’t matter enough to switch between models and harnesses that often.

(Despite that, I’m a novelty junky and I still oscillate between Claude Code, Cursor, and a couple of other approaches more often than I would like to quantify.)

Optimistic prediction: there will be a lull in frontier model capability race in the next 6–12 months. It will either correspond with IPO quiet periods or increasingly bad optics about the development of increasingly autonomous models. I’m optimistic about this because I think we could also use a breather in this domain.

2. Skills & prompting

Models and agents are good enough, in the late summer of 2026, to focus on customizing and leveraging skills instead of the model of the month. Bonus, skills are portable across agents, and models, sort of.

For most folks, those living in the middle of the bell-curve (definitely not the Amp folks, who seem to exist in the future), I think this means the action is now about telling the agent what the task is and how to figure out if it’s done or did a good job. Prompting and verification.

I’ve had a lot of luck lately with simplifying my instructions and moving more of the details into (Markdown) documents. In general, a lot of what I had in my CLAUDE.md ten months ago feels like over-steering now.

That said, moving the quasi-deterministic “verbs” of your interactions with coding agents to skills still makes a lot of sense. In particular, to socialize usage patterns with teammates or codify how specific tools, e.g., Xcode or CSS, are used.

Optimistic prediction: most currently-popular skills will be trained into the models. Instead of /grill-me about the new login screen or /fix the bug where the button disappears evaporate into garden-variety English.

3. The human factor

A couple of weekends ago, I was tinkering with doing Claude code via iPad and remote control. But I remembered the smarter move is to avoid tokens on evenings, and probably weekends too.😇

Lately, I’m finding that the drive for software factories and sprawling specs feels too much like the boring kind of software development work. Nudging and steering an agent, asking questions and getting answers in the form of working code, that’s more fun. It’s not as surprising and energizing and sometimes-frustrating as human collaboration. But it doesn’t feel like doing rote homework problem sets, which must count for something!

Optimistic prediction: the models whose output is most concise and least cliché could prevail over more capable but inscrutable ones.

4. Even more predictions

In the coming months:

• More hobbyist/self-employed/solo developers will utilize >1 agent harness monthly subscription. For bouncing back and forth between whichever model is in the out-performing at the time. Or if one prefers “the vibe” of one model as a collaborative partner while using the other as a coding agent. Sometimes, to get extra token usage at $40/month instead of going all the way to $100/month. More reasons to carry subscriptions to two or more models will probably come along as use-cases evolve.
• As more developers carry access to more models, skills and workflows will gain traction that use models in an “adversarial” mode. That is, model A reviews model B’s plan, model B review model A’s code, etc. Or, model A/B/C all try to find bugs/security issues in existing code in a competition. I’m not sure how this will pan out, but I’ve heard of people having good luck with this currently.
• One of the frontier model labs could go public. We get a better idea about the economics of their operation, what token subsidies actually look like in numbers, the actual economics of training and releasing a new model, and what the labs think their long-term play is currently. (Pessimistic: their filings and road-show are just “come with me, it will be good, but we’re doing it my way” like Facebook, and we don’t learn anything.)
• That developer with Pi kitted out just like they like it is the new co-worker who insists on vim (not NeoVim) or emacs with their dot files just like they like it. Same as it ever was.

It feels boring to write about agent coding so often. I try to keep it spaced out, even if it’s with modest vacation photos and ramblings. But, so much is changing and often interesting that it feels like missing out on the moment to develop one’s thoughts, in writing and publicly.

So, I regret to inform you, the discourse on AI will continue until the situation becomes less dynamic, fascinating, and rewarding to develop what I think of it by writing publicly.