University of Michigan professor Jason Mars is developing Jac, a Python-like programming language, and Jaseci, its runtime, to make software easier for AI agents to use. His central argument is that today’s languages and toolchains were designed for people, leaving agents to navigate a fragmented mix of technologies. He points to the reported $165,000 cost of converting the Bun framework from Zig to Rust with an AI model as an example of how much effort language translation can demand. Jac aims to consolidate that complexity: it supports Python and TypeScript ecosystems, compiles to Python bytecode, JavaScript, and native code, and can use packages from PyPI and npm. LLVM integration adds Rust-like ownership and borrow checking and access to C and C++ libraries; an “eject” feature can translate Jac code back into Python, TypeScript, or C for compliance. Java is not supported.
Jac also structures core data as graphs rather than conventional database tables or objects. Mars argues that agents can navigate graph relationships directly, avoiding token-intensive extraction and conversion, while “walkers” process data as they traverse it and can run in parallel. In a demonstration, an AI agent built a multi-stage, parallel tool-using clone of AI code-review service Greptile with Jac, while the same task using LangGraph produced a basic linear workflow. Mars takes this as evidence that the tools and abstractions given to agents shape the quality of their work. Jac has already been used in smaller commercial services, but its broader adoption remains unproven.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.