Thoughts on coding agents

Before December 2025, I was thinking about coding agents mostly as a question of autonomy, e.g. how far a model can go on its own before a human has to step in.

The models are being trained to produce trajectories that are effective for people who already know where they want to go. The value is not really the model guessing your intent. The value is the model compressing the distance between a clear intent and its execution. Once you see it this way, the interesting question is “how cheap is it now to get from a decision to a working artifact.”

That reframing changes who benefits. The people gaining productivity are not the ones handing an agent a vague goal and hoping. They are the ones with a strong internal model of the outcome like people who can steer, correct course, and recognize a good trajectory when they see one. The agent is an amplifier of clarity, not truly a substitute for it. If you know where you are going, it takes you there fast. If you don’t, it will happily take you nowhere.

There is a second thing we got wrong. We called them coding agents because we expected them to write software, and we quietly assumed that was the boundary of their usefulness. It turned out that generating code is not a narrow skill at all. Code is the most general interface we have for acting on the world. It can call any API, transform data, drive tools, and automate almost any process a person can describe. The moment a model could reliably turn intent into working code, it inherited the reach of everything code can touch.

So the coding agents quietly became a general-purpose harness for getting things done. The “coding” part was almost incidental. The real capability was the iteration: propose a trajectory, run it, observe what happened, correct. Keep doing it until you run out of tokens or achieve your goal.

Here is why this matters beyond individuals. For most of my life the real cost of building something that solves a hard problem inside a large organization was not always about execution. It was often coordination. Organizing people, aligning teams, and convincing enough stakeholders while code is being written. Execution overhead was almost always dwarfed by social overhead. You spent your best energy earning permission to start, and then do it again and again to be able to continue. Coding agents removes this traditional permission system. It benefits some and simultaneously creating unmanageable chaos for others.

Coding agents collapse the execution cost so aggressively that the bottleneck moves. An individual with goals can now go from intent to a working prototype without assembling a group first. This reduces to burden to show something but also enables new levels of cookie licking. It also creates chaos by making it too easy to fork an idea or a project when no one is in charge to make broad decisions for organizations. This puts traditional organizations in a tough spot: they either need to gate keep themselves or make their engineers find opportunities somewhere else. Neither of these options are good options.

This is the actual explanation for why things look so uneven right now. The agents remove the typical tax that large organizations quietly levy on individual conviction. If you are positioned well to leverage this, you may end up having a generational advantage to build early in fields where there is a lot of novel work to do. So, the overall impact of the coding agents seems to be a question about people (organizationally), and about how much easier it has become to act on without permission.

Finally, this is why I keep exploring what these models can and cannot do. It matters to me to understand what it means for all of us, not just for the work in front of me. The part that worries me is how fast the dynamics may be changing. Senior engineers no longer spend enough time educating the newcomers, because the amplifier rewards those who already know where they are going. But knowing where to go is exactly the capability that has to be learned by observing. If we stop passing it on, we risk raising a generation that can execute but cannot set the goal. In the long term, losing the ability to decide where to point all this leverage is a real risk to technological progress, and it is the thing I want us to keep in view.