MLIR-AIR agentic harness v0.1

The v0.1 agentic harness supports the AMD XDNA2 NPU compilation flow built on the MLIR-AIR compilation stack. It connects Codex or Claude Code to AIR implementations, the measured kernel registry and 17 NPU skills. The agent iterates through kernel development and model composition, using source inspection, compiler feedback, fixed correctness checks and production profiling.

Start the coding agent in the AIR repository and use $deploy-new-llm OWNER/MODEL in Codex or /deploy-new-llm OWNER/MODEL in Claude Code. Add precision, workload or acceptance requirements to the same request. The agent resolves a pinned task, implements the required model adaptation, then verifies and profiles it. Existing supported models can use reproduction mode.

The package quick start includes task preparation, a runnable example, prerequisites and host tests. The workflow explains the two loops. Client installation uses portable repository registrations pointing to one canonical skill payload. Both clients use the shared project instructions; the installation guide explains the Claude Code version and loading settings for this file.

Kernel reuse follows the declared arithmetic, layout and validation contract. The registry evidence interface adds opt-in verified selection while preserving legacy measured configurations. Model checks use complete finite captures before applying mutual top-5 token inclusion; measured performance retains the request boundaries and raw samples.

The harness extends From Human Guidance to Autonomy: Agent Skill System for End-to-End LLM Deployment on Spatial NPUs, presented at MLArchSys'26, co-located with ISCA 2026. See Used in / Cited in for the citation.