Presents SMART, a symbolic performance-modeling library that is fully regenerated by AI coding agents rather than incrementally patched, to escape perpetual refactoring caused by changing ML assumptions.
SMART is an AI-native machine-learning performance modeling library where the main branch contains almost no code, relying instead on a directed acyclic graph (DAG) of natural-language design docs as the durable artifact. This approach allows AI coding sub-agents to fully regenerate the implementation from the documentation on every version update, eliminating the technical debt and perpetual refactoring associated with incrementally patching aging frameworks.
Key features of this system include:
This method treats code as a regenerable build product, ensuring that the performance models remain accurate and up-to-date despite the rapid evolution of ML architectures and hardware systems.
The paper introduces SMART, an AI-native machine-learning performance tool built around the idea that a performance-modeling library should be continuously regenerated from its design documents rather than maintained as a long-lived, incrementally patched codebase. Rather than treating the implementation as the primary artifact, SMART treats its design docs—encoding the symbolic performance model, assumptions, interfaces, and invariants—as the canonical source of truth. AI coding agents then use those documents to regenerate the library, allowing the tool to track rapidly changing ML assumptions without accumulating the technical debt that typically comes from repeated manual refactoring.
The central contribution is a shift in maintenance strategy: instead of adapting a fixed codebase to new ML workload characteristics, hardware constraints, scheduling assumptions, or modeling abstractions, the system regenerates the code from updated specifications. This is especially relevant for ML performance modeling, where the “correct” model can change quickly as new architectures, accelerators, memory hierarchies, and execution patterns appear. By keeping the design docs expressive enough to capture the symbolic model, SMART aims to reduce the perpetual drift between the intended performance semantics and the actual implementation.
The work matters because it proposes a more sustainable path for building ML systems tooling in an era where both the models and the toolchains are evolving rapidly. If performance tools can be regenerated from well-structured design documents, teams may gain faster adaptation, better consistency, and lower long-term maintenance cost. More broadly, the paper illustrates an emerging AI-native software-engineering pattern: rather than using AI to patch existing code, use it to regenerate implementations from high-level, maintainable specifications, making documentation—not legacy code—the durable engineering asset.