Quasar: A New Programming Language Designed for LLM Code Actions

Large language models frequently rely on writing executable code to complete complex tasks—using loops and conditionals to reason through problems. Traditionally, this code is written in Python, a language where LLMs demonstrate strong capability. Yet Python's design makes it hard to add many features that would strengthen these code-driven interactions with external tools.

Separating Logic from Tool Calls#

A team of researchers proposes a cleaner architecture: isolate the internal program logic that drives decision-making from the external calls that interact with the outside world. This separation opens up room to implement new capabilities. Features can be added by (1) marking external calls with metadata about their effects, and (2) modifying how the internal execution engine tracks and respects that metadata.

Introducing Quasar#

The team developed Quasar, a novel programming language built on this principle. The language demonstrates its power through three practical enhancements to LLM code actions:

• Access Control with User Prompts: Built-in security that batches user queries, reducing unnecessary interruptions while maintaining control over sensitive operations.
• Automatic Parallelization: External function calls run concurrently where safe, cutting response latency.
• Conformal Prediction: Uncertainty quantification techniques that help mitigate hallucinations and improve reliability.

Why It Matters#

By decoupling program logic from tool interactions, Quasar allows language designers to add powerful features without fighting Python's existing constraints. This could make LLM-driven code agents more efficient, secure, and trustworthy for real-world applications.

The research was submitted to arXiv in June 2025, with an updated version in August 2026.