Nvidia's research lab released SoL-Pi, an open source extension for the Pi coding agent framework built to reduce token usage, model inference work and agent turns on long-running tasks, according to the project's GitHub repository.
The tool combines four mechanisms: fusing file edits with validation checks into single tool calls, converting repeated large outputs into compact retrievable handles, compressing diagnostic logs while preserving verifiable quotes as evidence, and dropping completed plan steps from active context once they're no longer needed.
SoL-Pi is built as a standalone add-on rather than a patch to Pi itself, opt-in by configuration, and designed to preserve evidence and respect Pi's existing runtime choices, the repository says. It is released under the MIT license and targets Pi version 0.85.1 and compatible releases.
The project grew out of Nvidia's internal research into making agent frameworks themselves more efficient, a step the team says comes before trying to scale up how many agent loops run at once.
For teams running coding agents at volume, the pitch is direct: the same task done with less redundant context sent back to the model on every turn, which matters for anyone paying by the token.