TIRx Harness is a compiler harness combining a minimal stable compiler foundation, a knowledge base, tools, and a benchmark server to help agents develop correct, fast GPU kernels.

TIRx Harness brings together:

• TIRx-lite for kernel authoring: a domain-specific language over the TIRx intermediate representation.
• Compiler analysis: check synchronization, memory races, and numerical behavior; inspect compiler output and generated GPU instructions.
• kcoral for remote execution: keep your agent on one machine and run GPU work on another.

Skills guide the agent in using them, while workload contracts define correctness and performance.

Install the released package from PyPI:

python -m pip install tirx-harness

To develop the harness, follow Build from source for the build prerequisites, repository checkout, and native submodule setup.

See the documentation for details:

• Installation: prerequisites, other installation methods, and agent skills
• Quick Start: optimize a kernel in your own project
• Optimization Runs: run a registered workload

The book Agentic GPU Programming for MLSys explains the design behind the harness and walks through an optimization workflow.

• Add a workload: define a new optimization task
• Contribute a kernel: publish a kernel produced by a run
• Report and fix bugs: reproduce and resolve a defect