Qualities of Great AI Coding Agents

I wrote about what makes a coding agent great back in January, a list of qualities with no real story behind any of them. Nine months later I have actual stories, because the problems I listed then were problems I was living through, and most of them have quietly gotten solved. This is the update, five qualities I actually hit friction on, and where each one stands as of September 2026.

context that survives across a session

The first problem was memory, and it's mostly solved now. I used to repeat the same preferences every time a session changed, no emojis in frontend design, no gradient backgrounds, no em dashes in copy, the same instructions typed out again and again because nothing carried over. Claude specifically remembers this now, project and user level memory that persists across sessions, so I say it once and it holds. That's not a small convenience, it's the difference between managing a tool and actually delegating to one.

asks instead of assumes

The second one might be the biggest shift I've seen in coding agents, period. Early tools either did exactly what you typed with no judgment of their own, or confidently overrode you with their own opinion instead. Claude's version of this is different, it gives a real recommendation, states its own read on the problem, and still folds my actual opinion into the solution instead of discarding it. That balance, a point of view that doesn't need to win, is a turning point, not an incremental feature.

knows what it doesn't know

This is the one I'm least sure is actually fixed. Earlier, on text-heavy pages, agents would just conjure information that wasn't there, or worse, generate a fake API key and hit a live endpoint as if the integration were real. I haven't watched this fail recently, but I also haven't stress-tested it hard enough to call it solved. This is the quality I'd still ask a founder building in this space the most pointed questions about.

works with what's actually there

Messy code used to be a wall, now it's just friction. Agents would flatly refuse to touch anything undocumented or poorly structured, which meant the exact code that needed the most help was the code they wouldn't go near. That's being handled well now, refusal has turned into actual engagement with the mess.

stays inside scope

The last one is about restraint, and it's the most underrated quality on this list. Agents used to hand back three things you didn't ask for alongside the one thing you did, which meant every request needed a review pass just to find what actually mattered. That's solved well now too, you get what you asked for, not a bonus tour of everything the agent thought you might also want.

conclusion

None of this was true nine months ago, and I don't think it'll stay this settled for long either. The list changes because the tools do, and I'd rather keep updating this honestly than pretend I called it all correctly back in January. Four of five, genuinely better. One, I'm still watching.