A six-part blog series: field notes from building Ignis, a Mojo-native agent harness that runs a MAX-served LLM in the same OS process as the control plane. Less “release notes,” more “what I found going into the innards of Mojo and MAX.”

New to Mojo or MAX? Parts 3 and 4 each open with a from-scratch, five-minute primer - enough to onboard someone who has never touched either before diving into the assessment. If you already know the tools, skip those intros and go straight to the details; the rest of the series doesn’t repeat them.

The series

  • Part 0 - The Curiosity and the Plan - the trailhead: what made me curious about Mojo and MAX, the ds4 spark, the in-process bet, and the M0→M5 plan I set out with.
  • Part 1 - What It Is, and How to Run It - the artifact and a functional user guide: the in-process design, the Engine trait everything hangs off, the architecture and control flow, extensibility, the examples, and how to build and run it.
  • Part 2 - What Was Achieved - a scorecard against the plan, headlined by durable cross-process session resume, with the numbers I measured and the ones I deliberately didn’t.
  • Part 3 - Mojo, the Language - a from-scratch primer, then an assessment of Mojo at 1.0 beta from building in it: where it matured, where it bites, and why batteries-not-included decides fit.
  • Part 4 - MAX, the Platform - a from-scratch primer, what MAX is and its place among NVIDIA’s stack, vLLM/SGLang, and the MLIR compilers; then the deep climax - compiling Ignis’s refund policy into the decoder two ways.
  • Part 5 - My Journey - the trail told as a trail, the verdict, and a genuine appreciation for what Modular built.

Back to the blog home.