NVIDIA’s SoL-Pi cuts token use in long coding-agent runs

NVIDIA researchers have released SoL-Pi, an open-source extension for the Pi coding-agent harness that reduces repeated work during long tasks. Its four opt-in mechanisms combine predictable actions, replace replayed output with handles, condense verified evidence, and compact completed context. The team selected those mechanisms through a scaled auto-research pipeline in which agents proposed, implemented, tested, and reviewed harness changes.

A coding-agent harness coordinates the model, conversation history, editing tools, shell commands, and returned observations. As a task stretches across hours, the harness may repeatedly send old context, ingest large tool outputs, write caches, and request extra model turns. Those costs accumulate even when the underlying model and task remain unchanged.

NVIDIA connects this efficiency problem to recursive self-improvement, or RSI, in which AI systems help modify the software and workflows used to develop subsequent AI systems. SoL-Pi tests whether AI-assisted harness research can make those loops cheaper while maintaining a declared capability floor.

Four controls for recurring waste

SoL-Pi installs through npm and runs on an unmodified release of Pi. The extension leaves every mechanism disabled until the developer enables it in configuration, allowing teams to evaluate each change separately.

• Action Fusion lets a file edit or write invoke its follow-up validation command in the same tool call. Analysis of prior trajectories found that edit or write transitions followed by commands represented 12.3% of cross-turn transitions, with Bash accounting for 85.1% of those next actions.
• ObservationPack replaces repeated large text results with stable handles and exact paged recall. It archives the original payload locally, allowing the agent to retrieve specific pages without replaying the entire result.
• Evidence-Preserving Reducer converts long diagnostic logs into compact sets of verbatim evidence. A cheaper model reads the full log first, and every retained quotation must match the archived source before the primary model receives it.