Pi vs OpenCode: local coding on Strix Halo and R9700

This AMD-sponsored episode compares Pi and OpenCode as coding agents backed by a local LLM. The speaker explains how an agent harness manages context, exposes file and shell tools, and saves sessions. He argues that Pi's smaller prompts suit hardware where processing long context takes more time.

Both agents tackle a ROCm toolbox update from 7.2.2 to 7.2.3, including related scripts and GitHub workflows. The demo uses Qwen 3.6 with an 8-bit Unsloth quantization on Strix Halo through llama.cpp. The speaker reports about 25,000 tokens for Pi and 39,000 for OpenCode, which takes longer in this example. OpenCode provides built-in permissions and plan/build modes, though the model writes files while still in plan mode during the demonstration.

The setup discussion covers a Framework Desktop with 128 GB of unified memory and a Radeon R9700 AI PRO with 32 GB of VRAM. It also shows server access through SSH port forwarding and Pi sandbox restrictions for files and network domains.

An adapted SWE-bench mini evaluation tests Pi on 50 tasks, with Gemini 3.1 Pro judging solutions. The speaker reports shorter task durations with multi-token prediction in llama.cpp. These results have limits: possible training-data contamination, variation between runs, and autonomous execution without internet access or human feedback. The episode also explains prompt injection risks and ways to limit an agent's access.