The Harness Is the Runtime: Pi's Sub-1,000-Token System Prompt and 4 Tools Beat Claude Code and Codex on Databricks' Own Codebase at 2.08x Lower Cost
An August 4, 2026 essay argues minimalism is the whole advantage for coding harnesses, citing Databricks' independent benchmark on its multi-million-line codebase: paired with Opus 4.8 at xhigh effort, Pi — four built-in tools and a system prompt under 1,000 tokens — took the highest overall pass rate, and at high effort was 2.08x cheaper at comparable quality (85% vs 87%) while sending roughly 3x less context per turn. The claimed mechanism is that bloated tool surfaces bury instructions in the hierarchy and inflate per-turn context. The post also surfaces Shopify's pi-autoresearch, an autonomous optimization loop built as a Pi extension that runs experiments and reverts regressions, reportedly yielding 300x faster unit tests and 20% faster React components.
Source
↳ Follow the thread