Software is Mostly All You Need
Neural Networks at Buildtime, Software at Runtime
Over the last 6 months and the last 6 weeks in particular, AI coding agents have shown to be incredibly capable at writing software. Tasks that traditionally required weeks of human labor can now be done in days if not hours. Even more incredibly, software systems that are designed from the start to harness AI coding agents exhibit many of the characteristics of the neural nets that were integral to their creation in the first place. These AI-native software systems are learned, not designed. Code is the policy, deployment is the episode, and the bug report is the reward signal - well-architected coding agents can drive this loop with little human intervention. Unlike traditional reinforcement learning architectures, they are encoded in CPU instruction sets instead of neural network weights, but they are learned just the same.
The success of coding agents and the software systems built thereon carry lessons about where to apply AI agents in general as well. Coding, like many other creative tasks, requires judgment. How best to implement some function with input A and output B; how to name some variable; whether to share some function or implement a new version; etc. Neural networks excel at judgment (more on why below). Yet many of the agentic deployments we are seeing in the wild are against tasks that can be fully specified as explicit instructions. Of course, traditional software excels at executing explicit instructions. Any programming language can be executed on today’s machinery at billions of instructions per second.
Coding agents get this exactly right, since by definition they are making a series of judgments when writing code at buildtime and leaving the execution of such code to machines operating at runtime. The best performing architectures follow suit, delegating judgment to neural networks and execution to traditional software, even when the executable artifacts are produced entirely by AI.
Some Agents in Practice #
Many agentic AI projects are failing — agentic drift, opaque debugging, brittle autonomy. Meanwhile, Claude Code has driven significant productivity gains by doing something different: it writes code that humans review and deploy, producing artifacts that are durable, version-controlled, and deterministic.
These failures and successes reflect a fundamental architectural difference.
Judgment and Execution Historically #
Humans have historically done two different types of jobs for different reasons, and AI changes each differently.
... continue reading