Modular says durable compiler infrastructure matters more in the agentic era

Agents increase the need for dependable compilers

Modular’s Michael Dunn-O’Connor writes that accelerator code is multiplying faster than the shared tooling around it. Tuesday’s keynote described agentic development as “the elephant in the room”, as teams can rapidly generate code optimized for specific devices.

The report compares agents with Daniel Kahneman’s System 1 thinking, which is fast and intuitive, and compilers with System 2 thinking, which is deliberate and rule-bound. Kristof Beyls argued that agents do not make compilers less relevant. Instead, the dependable and verifiable layer becomes more important because generated code needs reliable systems behind it.

Beyls identified three requirements for durable open source projects: communicating intent, maintaining cognition and supporting contributors. The discussion also highlighted “contributor poker”, the practice of investing maintainer effort in a new contributor even when an initial pull request needs substantial work.

Shared tooling could reduce duplicated LLVM maintenance

A Wednesday keynote from Google’s Nick Desaulniers, Anshul Nigham and Jordan Rupprecht examined the challenges of maintaining a tip-of-tree fork of LLVM. The maintainers found that delaying an update can create further delays as breaking changes compound.

They also found that teams were applying LLVM and MLIR updates independently across the same monorepo, duplicating maintenance work. Their proposed solution was better shared tooling to make updates more efficient and coordinated.

For Modular, the issue reflects a broader tension. One-off solutions can be produced quickly, but drifting from common architecture and tooling can leave software harder to understand and maintain.

DSL growth creates an opportunity for Mojo and MAX

As new accelerator generations arrive alongside older hardware, developers need to support multiple architectures while preserving performance and readable code. Modular says this is driving the creation of more Domain-Specific Languages, or DSLs.

NVIDIA’s BaseDSL project, presented by Maksim Levental, Guray Ozen and Amir Mohammad Tavakkoli, is a platform for building custom Python DSLs on MLIR through plugins. The approach can reduce duplicated effort when teams create DSLs, but the report says teams may still end up with incompatible languages and separate maintenance burdens. BaseDSL is internal to NVIDIA for now.

Modular presents Mojo as an alternative shared layer. The company says Mojo is a full language that lowers through MLIR, with readable syntax and metaprogramming for control and performance across accelerators. MAX adds accelerator support, using Mojo as its programming interface, so hardware support can be built in MAX and optimized through one language rather than separate team-specific DSL stacks.

Open source communities still need human collaboration

The report describes frustration among LLVM maintainers over large volumes of drive-by AI-generated pull requests, particularly when no human developer remains available to respond to feedback and complete the work. Modular says the challenge applies both to external contributions and to the maintenance of critical software.

The company argues that shared standards and abstractions must be accompanied by people willing to collaborate, develop a common vision and mentor contributors. It points to the Mojo contributor guide and Modular AI Tool Use Policy, inspired by the LLVM AI Tool Use Policy, as examples of practices supporting that community.

For users and developers, the message is that faster code generation does not remove the need for comprehensible projects and coordinated infrastructure. Modular sees the combination of Mojo, MAX and community practices as a way to reduce duplicated accelerator work while keeping software aligned across teams and hardware.