⚡ SoL-Pi: Scaling Auto-Research Loops for Efficient Agent Harnesses
[!NOTE] This repository contains the open-source version of SoL-Pi, a standalone extension for Pi. It is not an official distribution of Pi.
💡 TL;DR
Spend less without making the agent do less useful work.
SoL-Pi is a standalone extension for Pi that packages four reusable efficiency mechanisms discovered through scaled auto-research loops. It reduces repeated model turns, context replay, oversized observations, and unnecessary long-log reading while preserving the work and evidence an agent needs to finish a task.
SoL-Pi installs on top of an unmodified Pi release. Every mechanism is opt-in and disabled by default.
Introduction
Long-running coding agents accumulate repeated work. A file edit is often followed by a predictable validation command. Large tool results are replayed long after their first use. Completed subtasks remain in active context, and a frontier model may spend a full request reading a log when only a few lines affect the next decision.
SoL-Pi grew out of a broader question from our auto-research work: before scaling agent loops, can agents first make the harness itself more efficient? The search focused on constrained efficiency: reducing token traffic, inference work, and agent turns without stopping early, skipping verification, or hiding evidence.
The standalone release contains four mechanisms that survived that process. They operate at different parts of the harness and compose through Pi's public extension APIs.
What SoL-Pi Adds
| Area | Mechanism | What changes |
|---|---|---|
| Tools | Action Fusion | An edit or write can run its follow-up validation command in the same tool call. |
| Observations | ObservationPack | Repeated large text results become stable handles with exact paged recall. |
| Delegation | Evidence-Preserving Reducer | Long diagnostic logs become compact receipts only when every retained quotation matches the archived source. |
| Context | Online Context Compact | Completed plan steps become candidate points for Pi's native compaction, subject to economic and window-pressure checks; after a successful compaction, Pi continues the task in a new turn. |
The mechanisms share four rules:
- No Pi patches. SoL-Pi imports public Pi APIs and does not vendor the Pi source tree.
- Explicit opt-in. A missing configuration leaves every mechanism disabled.
- Preserve evidence. Original observations remain available locally, and reducer failures leave the original result unchanged.
- Use Pi's runtime choices. Authentication, provider URLs, the main model, and shell behavior remain under Pi's control.
Technical Details and Core Insights
Read the SoL-Pi blog for a deeper look at the technical details, design rationale, and core insights behind SoL-Pi, including how auto-research led to the four efficiency mechanisms and how they work.
Paper
Read our paper: [SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness](https://arxiv.org/abs/2609.