Earendil has set out the case for Pi, a coding-agent harness built around a deliberately small default toolset and an extension system. The company says Pi begins with four tools, while its system prompt and tool definitions together occupy fewer than 1,000 tokens.
The approach aims to reduce the instructions and repeated context surrounding an AI model. Instead of including many workflow features by default, Pi lets users add capabilities for their own needs. Earendil calls this “context discipline” and argues that a smaller working set can lower costs and reduce the number of model runs needed for a task.
The company points to a Databricks study of coding agents on tasks drawn from its multi-million-line codebase. According to Earendil’s summary, the study separated the underlying model from the harness used to operate it and found that the harness could materially change costs and results. Earendil reports that Pi paired with Opus 4.8 at the study’s “xhigh” setting achieved the highest overall pass rate while costing substantially less than Claude Code and Codex.
Earendil also says the Databricks work found more than a twofold difference in cost per task in some comparisons using the same model and thinking effort, even where quality was unchanged. Pi sent about one-third as much context per turn, according to the account. These figures are presented in Earendil’s advocacy for its product and were not independently substantiated in the supplied evidence.
A second example concerns Shopify’s pi-autoresearch extension. Earendil describes it as an autonomous optimization loop that conducts measurable experiments, rejects regressions and retains improvements. The company says Shopify reported cases in which unit tests became 300 times faster, React component mounting improved by 20%, and build times and pnpm performance also improved. Those are reported case outcomes, not a guarantee for other codebases.
For Earendil, the example is meant to show that a minimal starting point need not prevent specialized automation. Pi can inspect its extension documentation and help create a workflow, allowing complexity to be added selectively rather than bundled into every session. The company makes a similar argument for local models, where smaller context windows and slower prompt prefilling can make a stable, compact prefix particularly useful.
The broader claim is that coding-agent economics depend on the entire run rather than token price alone. A more capable model may still be cheaper if the harness helps it finish in fewer steps, while a lighter model can cost more if it requires repeated turns. Earendil’s evidence frames Pi as one possible implementation of that principle. Independent comparisons across models, repositories and task types would be needed to establish how consistently its advantages hold.
