NVlabs/SoL-Pi
SoL-Pi: Scaling Auto-Research Loops for Efficient Agent Harnesses
What it solves
SoL-Pi addresses the inefficiency of long-running coding agents that accumulate repeated work, replay large tool results unnecessarily, and waste tokens reading long logs or maintaining bloated contexts. It aims to reduce token traffic, inference costs, and the number of agent turns without sacrificing the quality of work or the evidence needed to complete a task.
How it works
It functions as a standalone extension for the Pi coding agent, implementing four opt-in efficiency mechanisms:
- Action Fusion: Combines a write or edit action with its subsequent validation command into a single tool call to reduce turns.
- ObservationPack: Converts repeated large text results into stable handles with paged recall, preventing the same large data from being replayed multiple times.
- Evidence-Preserving Reducer: Compresses long diagnostic logs into compact receipts, ensuring that every retained quote matches the original source.
- Online Context Compact: Identifies completed plan steps as candidates for context compaction based on economic and window-pressure checks, starting a new turn after successful compaction.
Who it’s for
It is designed for developers and researchers using the Pi coding agent who want to lower their operational costs and improve agent efficiency during complex, long-term coding tasks.
Highlights
- Non-intrusive: Installs as an extension without requiring patches to the core Pi source code.
- Evidence-focused: Ensures original observations remain available and prevents data loss during reduction.
- Opt-in: All efficiency mechanisms are disabled by default and must be explicitly enabled via configuration.
- Pi-compatible: Leverages Pi's public extension APIs and maintains Pi's existing authentication and provider settings.
Related
- Project
- Project
- Project
- Project