Pi, Minimal and Performant

Everyone building AI coding agents seems to be racing toward the same destination: bigger prompts, more tools, thicker layers of orchestration. Pi went the other way. It ships with exactly four tools and a system prompt that clocks in under 1,000 tokens, betting that most coding work doesn't need a kitchen sink of features bolted on by default. That bet is now backed by real numbers from two companies that actually tested it against the competition.

Databricks ran the experiment properly. Instead of leaning on public benchmarks that everyone has by now optimized against, they built their own test suite from tasks their engineers handle daily inside a codebase with millions of lines. Paired with Opus 4.8 running at high reasoning effort, Pi posted the best pass rate of any harness tested, and did it for meaningfully less money than Claude Code or Codex. The kicker: same model, same thinking budget, but cost varied by more than 2x depending on which harness carried it. Pi sent roughly three times less context per turn and needed fewer runs to finish the job. Databricks called it out directly, quality didn't move, but the price tag did, purely based on how disciplined the harness was about what it fed the model.

Shopify's story is different but points the same direction. Engineer David Cortés basically asked Pi to build its own extension, something called pi-autoresearch, and Pi pulled it off by reading its own documentation and constructing the workflow from scratch. That extension turned into an internal optimization loop that runs experiments, throws out regressions, and keeps tightening code automatically. The results Shopify reported are not subtle: unit tests running 300 times faster in some cases, React component mounting sped up by 20%, faster builds across several projects, even gains in pnpm performance. None of that came pre-installed. Pi just made it easy enough to build that an engineer could ask for it in plain language and get a working tool.

The deeper argument here is about timing. A year ago, harnesses built natively around a specific model had a real edge because models were tuned to expect that structure. That edge is fading fast. Anthropic trimmed Claude Code's system prompt by 80% recently, which is a tell that frontier models no longer need hand-holding to operate a terminal-style environment. What matters now is whether a harness wastes context or keeps things clean, and that's exactly the lane Pi has picked. Because it's built to be extended rather than pre-loaded, teams add complexity only where it actually pays off, instead of inheriting someone else's guess at what every user might need.

There's also a local-model angle worth flagging. Smaller context windows and slow prefill times make context discipline even more valuable on local models, since a stable prompt prefix avoids minute-long re-processing every time something changes. Pi's minimal, unchanging default setup fits that constraint almost by accident, which suggests its advantages might compound as more coding work shifts off cloud-hosted frontier models and onto local hardware.