NVIDIA Researchers Present SoL-Pi, Reporting Up to 49% Less Token Traffic

A research team affiliated with NVIDIA, NTU and MIT has presented SoL-Pi, a set of changes to a coding-agent harness designed to make long agent runs use less context and fewer tokens. On a 51-task evaluation, the authors report 44.7% to 49.0% less recorded token traffic than Pi while achieving comparable performance across two tested models.

The target is a familiar cost of long coding sessions: an agent moves through reasoning, tool calls and feedback while carrying forward a growing record of prior work. A harness coordinates those actions, the conversation history and returned information. SoL-Pi works at that layer, seeking to reduce the information and steps an agent repeatedly processes.

A lower-cost run, by the authors’ measure

The paper estimates hourly savings of $8.75 to $13.50 against native Codex and Claude Code harnesses, and $4.36 to $5.71 against Pi. Those are benchmark-based estimates, not a claim that every team will see the same savings in its own codebase and tool chain.

Four mechanisms emerged from the search

The team developed the system through auto-research loops at the harness layer. It explored candidates broadly, then refined the ones that survived across diverse environments. The final package covers action execution, context compaction, observation handling and delegated reading.

• Action Fusion changes how an agent executes actions.
• Online Context Compact focuses on shrinking carried context.
• ObservationPack changes how the agent handles observations.
• Evidence-Preserving Reducer addresses delegated reading.

The researchers say the retained mechanisms can transfer beyond the settings that produced them because the search covered numerous repository-derived and verifier-driven environments. That proposition is consequential: it suggests harness improvements may be reusable rather than tuned to one task.

The next move is testing the harness under real workloads

SoL-Pi’s result centers on a practical proposition: a coding agent may become materially cheaper when its harness stops repeatedly carrying unnecessary context through a task. EdgeBench provides an initial test across 51 tasks. The important next question is whether the same balance of lower token use and comparable task performance holds in longer, messier software work.