From AI demos to dependable work

That's when NextEdit suggestions started working with some real accuracy. Then we went to chat, then we went to actions, and now to full autonomous agents. And then the autonomous agents can be both foreground, background, in the cloud or local. So that's all the form factors that exist today when you're coding. And interestingly, if you look at it, you use all of them. It's not like there's only one form factor. So that's, I think, probably one of the other lessons. So for example, when I'm in a CLI, you go a foreground agent, background agent, and then just literally go edit in VS code right there, all happening in parallel. So that sort of shows how these form factors even compose. So then you bring that to knowledge work, to your point.

I think it happened with Opus 4.5. Opus 4.5. That was an inflection point when models were started being amazingly good at orchestration and reasoning and tool calls. And Cloud Code brought in this new idea in AI that everything can happen inside a sandbox, a console, a terminal with access to tools where tools are just command line tools. Yeah. They don't even need to have graphical user interface. So when you did that and when you organize around files and subagents and skills and CLIs, the models started becoming very good at handling the context. So the context window no longer became a problem. It just put whatever necessary into the context whenever it wanted to and dumped them away when it wanted to. And that made it suddenly so good at doing very long orchestration tasks.

I think it happened with Opus 4.5. Opus 4.5. That was an inflection point when models were started being amazingly good at orchestration and reasoning and tool calls. And Cloud Code brought in this new idea in AI that everything can happen inside a sandbox, a console, a terminal with access to tools where tools are just command line tools. Yeah. They don't even need to have graphical user interface. So when you did that and when you organize around files and subagents and skills and CLIs, the models started becoming very good at handling the context. So the context window no longer became a problem. It just put whatever necessary into the context whenever it wanted to and dumped them away when it wanted to. And that made it suddenly so good at doing very long orchestration tasks.

There was kind of like a perfect storm that caused the AI to do something and the agent to do something that didn't quite understand what it was doing. I think that if there is a systemic problem here rather than just kind of like a random edge case, it's that AI still doesn't know what it doesn't know. You know, like a human would stop before deleting a production database and just say, oh, I'm about to do something like really serious, really destructive. Am I sure I want to do this? You know, and a human would have stopped and said, oh, wait a second, like I need to be more confident in what I'm doing before I take that action. And AI still has this issue where, again, it can be kind of overconfident.

So AI is a lot harder to implement than anybody expected. For sure. You can do an agent pretty straightforward. You can prompt away, we cheated on tests, we cheated at work, did projects, improved our productivity 100x, right? All easy-peasy. We just assumed at the enterprise it would be just as easy. It's hard and it's terrifying and not terrifying for employees because Dario and everybody is saying 50% of white-collar people are going to lose their jobs. Here we are two years later, they said within two years and employment is still growing, people are hiring, we need more AI literate people, right? So CEOs have no clue what is going on, none whatsoever, and that's not going to change, right? And so trying to explain to them that you need to add a harness to all the, there's no chance, right?

That's when NextEdit suggestions started working with some real accuracy. Then we went to chat, then we went to actions, and now to full autonomous agents. And then the autonomous agents can be both foreground, background, in the cloud or local. So that's all the form factors that exist today when you're coding. And interestingly, if you look at it, you use all of them. It's not like there's only one form factor. So that's, I think, probably one of the other lessons. So for example, when I'm in a CLI, you go a foreground agent, background agent, and then just literally go edit in VS code right there, all happening in parallel. So that sort of shows how these form factors even compose. So then you bring that to knowledge work, to your point.