Most of an agent codebase is not the agent
The article argues that the majority of an agentic AI codebase consists of infrastructure, tool definitions, and integration logic rather than the agent's core reasoning. It highlights how frameworks like Jaseci's Jac language aim to reduce this overhead by focusing on the agent's decision-making logic itself.
Background
- The post critiques how many AI agent frameworks (LangChain, AutoGPT, etc.) conflate agent logic with tool definitions, memory systems, and model configurations, creating spaghetti code that's hard to maintain or debug.
- It introduces Jaseci's JAC language, a domain-specific language designed to declaratively separate agent behavior from the surrounding plumbing (tools, APIs, state management), so the "what the agent does" isn't buried in infrastructure code.
- Jason Mars (Jaseci's creator) previously led work on Jaseci, an end-to-end stack for building conversational AI; JAC is its programming language.
- The core argument: in most agent codebases, the actual agent logic is a tiny fraction of the total code—the rest is scaffolding. The article proposes using a DSL like JAC to invert that ratio.