TIRx Harness turns AI agents into GPU kernel engineers, with up to 6.84x speedups
JiaZhihao · x · 2026-09-30
Hongyi Jin's team released TIRx Harness, an open compiler harness that rethinks the compiler stack around AI agents, giving them an environment to explore, debug and optimize GPU kernels.
- Components: a minimal stable low-level compiler foundation, domain-specific compiler analysis for debugging guidance, a kernel zoo of 60+ kernels with hardware specs, and remote GPU evaluation for reliable benchmarks and profiling.
- Results: agent-built Kimi Delta Attention kernels achieved geometric-mean speedups of 2.94x over FlashKDA (forward) and 6.84x over FLA (backward).
- The team's bet: the next leap in agentic GPU programming comes from engineering the environment agents optimize in.
More from coding & agent
- Codex desktop Linux hang bug fixed in latest 26.928.20755 release — cedric_chee · 2026-09-30
- Hands-On Comparison of Top Gen AI Frameworks for Go in 2026: Genkit, Eino, ADK Go and More — rseroter · 2026-09-30
- mitsuhiko: use any llama.cpp model with Pi as a discount classifier — mitsuhiko · 2026-09-30
- Obsidian Mind: 4.7k-star project gives Claude Code, Codex and Gemini agents persistent memory — tom_doerr · 2026-09-30
- Ambion 0.4.0 Released, Focusing on Simplified Core Abstractions — andreisavu · 2026-09-30
- Upcoming talk: 'Spring AI: There and Back Again' on building AI apps with Spring — therealdanvega · 2026-09-30