HWBRIEF
Chris Lattner built LLVM, the open source compiler toolchain, and now leads advanced AI software at Qualcomm. He says coding agents help with narrow jobs like making a chip's code run faster, not with the software product other developers depend on.
Chris Lattner built LLVM and Clang, the open-source compiler toolchain used across the industry, and founded Modular, the AI software startup Qualcomm completed acquiring in July. He is now executive vice president of advanced AI software and platforms at Qualcomm, and he has a falsifiable test for when agentic code generation actually helps.
Agents help when the job is "well-defined and limited in scope," he told EE Times. The example: "make this kernel go fast," where kernel means a narrow performance routine on a chip, not an operating-system kernel. Work that can be "quantified, measured, and validated" is in scope.
The hard line is bigger work. "Bigger projects require product design and software engineering and can't be delegated to agents," Lattner said. Agents "make tons of throwaway code." His rule: agents help where success can be measured and validated; they do not help where the deliverable is a software product other developers will depend on.
Lattner hears both sides of that claim. Some hardware companies say they "don't believe in software anymore," he said, "they're definitely using LLVM and still investing in compilers, but saying they're not." And "various people use [agentic code generation] as a talking point, particularly when their software is known to be problematic."
Meanwhile, the LLVM project, which Lattner says is "under assault" from AI-generated code, now has its own AI tool use policy requiring a human in the loop and labeling of tool-generated contributions.