Auto-Syntactic Model: AI agents that live in the type system and own their language

Max von Hippel proposes a radical rethinking of AI coding agents: the Auto-Syntactic Model (ASM). In ASM, the agent does not exist as a separate program but lives inside the type system of the language it is programming in. Moreover, the agent has the ability to modify that language (its own ‘harness’) directly. This blurs the line between the builder and the built, making the agent part of the very software it produces. Von Hippel argues this is necessary because current state-of-the-art agents (Claude Code, Codex, Slate) have fundamental flaws that model improvements alone won't fix—namely, poor editing and search capabilities.

Von Hippel believes an ASM solves editing by giving the agent unique typed-object tags; it can then index into and edit specific objects instead of performing full-file rewrites, drastically reducing context size and error. For search, the agent can spawn sub-agents that live in the same type-system fabric, avoiding the need to load an entire codebase into context. The post also explores tangents: ASM could enable formal‑methods-based AI safety, and von Hippel shares a minimal weekend prototype built with Fable. Though still hand-wavy, the ideas challenge the conventional separation between tools and code.