Explore AI-agent-assisted translation pipeline (inspired by Modular's Mojo approach)

[Tokarzewski]

Context

Modular published a writeup on how they translate Python/CUDA/SYCL/C++ code to Mojo using general-purpose AI coding agents (Claude, etc.) augmented with a lightweight "skills" system, instead of a purely rule-based transpiler.

Source: https://www.modular.com/blog/translating-to-mojo-via-ai-agents

Key ideas from the article

• Skill-injected idioms instead of hardcoded rule tables: target-language conventions/gotchas (mojo-syntax, mojo-gpu-fundamentals, etc.) are packaged as retrievable docs the agent loads on demand (npx skills add modular/skills), compensating for the target language being too new/low-resource for the LLM to know natively.
• Compile-error-driven repair loop: the agent translates → compiles → runs against a reference/test harness → reads compiler errors and numerical diffs → self-corrects. Strict, typed target languages give much better repair signal than loose ones.
• Structural/idiomatic remapping over literal translation: for constructs with no 1:1 equivalent (e.g. CUDA warp-level tricks → portable shared-memory tree reduction), the agent restructures to the target's idiomatic form rather than doing line-by-line translation.
• Numerical/behavioral equivalence as the correctness oracle: when exact output isn't feasible (floating-point kernels, reordered reductions), validate via numerical drift (e.g. softmax row-sum ≈ 1.0 within float epsilon) and benchmark comparisons rather than only pass/fail unit tests.
• Reference implementations + published test harnesses: ground truth to diff against, usable by rule-based or agentic translation alike.

Proposed action items

• Evaluate whether a "skills"-style idiom-injection approach could reduce hardcoded pattern tables in our transpiler passes
• Investigate piping target-compiler errors back into an LLM-assisted repair loop for cases our static rules don't cover
• Identify constructs in our supported source/target languages that have no 1:1 mapping and would benefit from idiomatic restructuring instead of literal translation
• Consider numerical/behavioral equivalence checks as a fallback oracle where exact output matching isn't feasible
• Audit our test harnesses/reference implementations for reuse as ground truth in an agent-assisted pipeline