KARABASAN

Compil'agent reframes compilation as an agentic problem. It's a drop-in replacement for `torch.compile` and `@triton.jit`, and the main mechanisms are backend-agnostic. In just three lines of code, users get a faster, correctness-validated callable. What's distinctive is where the agent searches. Conventional autotuners sweep user-facing knobs like BLOCK_SIZE or num_warps. Compilagent goes a layer deeper, into the compiler's own decisions: Triton's MLIR pass pipeline, Inductor's scheduler, FX graph rewrites, or even lowering registry overrides. The agent reads the actual IR, forms hypotheses tied to specific evidence in those artifacts, proposes multi-knob candidate batches, benchmarks each, and reflects across the batch before the next round. The interesting bet Compil'agent makes isn't just that LLMs can do compiler search. This points to a broader shift in how compilers are built. The field has long shipped fixed pipelines tuned by hand-written heuristics that must work passably across every workload. Compil'agent treats the pipeline as a search space and the compiler's own decisions as a tool surface, and in the future, enables the agent to become the compiler itself.