Your software should build itself

I’ve recently decided that the distinction between the agent which builds your software and the software it builds is nonsensical and antiquated. In the future, software will build itself, and this means the agent is part of the very fabric of the software it’s building. To make this idea more formal, I propose two (I believe, novel) ideas:

Your agent should live inside the type system of the language it’s programming in.

Your agent should own (be able to edit) the language it’s programming in.

I will refer to a system which satisfies both attributes as an Auto-Syntactic Model (ASM).

The traditional distinction between a model and an agent is that a model is a single-step function that maps from text to text, while an agent is a model, with tool-use, in a while loop. [1] In an ASM, the agent writes code in a programming language, L, the syntax of which is defined locally (for example, using Racket). The agent has a modify tool by which it can modify L directly. Moreover, the agent does not exist as a piece of code distinct from the syntax and semantics of L; when it modifies L it is modifying its own harness. This is because the agent is implemented directly within the type system of its own language.

ASMs, I believe, solve several very significant problems with current SOTA programming agents. I enumerate these problems and explain why they necessitate ideas 1 and/or 2 in the next section, titled, Your agent sucks. Then, in the subsequent section, PL Supremacy, I enumerate a few compelling examples of really cool things you can do using ideas 1 and 2. After this I explore two interesting tangents. First, Tangent 1: AI Safety, explores how this new approach to agent architecture could allow for new and interesting AI safety mechanisms by way of formal methods. The second, Tangent 2: My Mediocre Prototype, briefly describes a very, very minimal prototype which I vibe-coded over the weekend using Fable. Finally, I conclude in Conclusion.

Note: This is ongoing r…