NVIDIA-Led SoL-Pi Breakthrough: AI Agents Now Operate with Up to 49% Less Token Traffic

Revolutionizing AI Coding Agent Efficiency with SoL-Pi

AI coding agents have become increasingly powerful, capable of running for extended periods to complete complex development tasks. However, this often translates to significant token consumption and associated operational costs. Addressing this challenge, a research team comprising experts from NVIDIA, Nanyang Technological University (NTU), and MIT has introduced SoL-Pi, an advanced set of four efficiency mechanisms specifically for the open-source Pi coding agent.

SoL-Pi stands out by implementing these mechanisms, which were discovered through an automated research process leveraging AI at the agent's harness layer. Comprehensive evaluations on the 51-task EdgeBench benchmark demonstrate SoL-Pi's remarkable ability to cut recorded token traffic by 44.7% to 49.0% compared to the baseline Pi agent. Concurrently, API costs are reduced by approximately 33%, all while retaining nearly 94% of Pi’s performance scores on both GPT-5.6 Sol and Opus 5 models.

Strategic Focus on the Harness Layer

Traditionally, efforts to improve AI efficiency have concentrated on reducing the cost per token through methods like faster computational kernels, model quantization, or employing less expensive models. SoL-Pi, however, adopts a distinct strategy by aiming to decrease the total number of tokens a task consumes. The chosen target for this optimization is the harness layer—the critical component responsible for managing tool interactions, contextual information, observations, and task delegation within the agent's operation.

Manually optimizing the harness is a time-intensive process, complicated by its interconnected parts where changes in one area can inadvertently shift costs elsewhere. While systems like Meta-Harness automate some of this work, previous research indicated that such evolved harnesses could overfit specific search tasks, yielding limited improvements on unseen problems.

The Auto-Research Methodology

The development of SoL-Pi's efficiency mechanisms involved a sophisticated auto-research loop. A dedicated research AI meticulously observed execution traces generated by a separate agent running the base Pi configuration. Based on these observations, the AI proposed novel modifications to the harness layer, which were then rigorously tested.

This extensive search covered an impressive scope:

• 152 distinct proposed directions, categorized across six families: context management, progress tracking, tool utilization, delegation strategies, prompt formulation and policy, and improvement/evaluation methodologies.
• 535 executable environments, including 495 derived from real-world GitHub issue-pull request pairs and 40 synthetic tasks featuring executable verifiers.
• Over 3,000 runs, resulting in more than 60,000 agent-environment interactions.

Each search instance functioned as an independent, isolated loop, adhering to an autoresearch cycle extended with a Ralph Loop implementation step and an independent review process. Crucially, acceptance rules for candidate mechanisms were predetermined and immutable by the optimizer, requiring adherence to capability metrics within defined tolerances and demonstrating improvement in at least one efficiency metric. The EdgeBench dataset was entirely held out for final, unbiased evaluation.

The Four Proven Mechanisms

The rigorous auto-research process identified four key mechanisms that ultimately constitute SoL-Pi:

• Action Fusion: The base Pi agent frequently performs a file edit followed by a separate command for testing or building. Action Fusion consolidates these into a single tool request, returning both outcomes in one observation, thereby eliminating an entire model round trip.
• Online Context Compact: This mechanism monitors plan steps via update_plan. Upon a step's completion, the harness estimates remaining requests. It then weighs projected input savings against the cost of rewriting the prompt cache, invoking Pi’s native compaction when advantageous or when the context window limit is approached.
• ObservationPack: For tool outputs exceeding 10 KiB, this system archives the full data locally. It transmits the complete output for the subsequent two provider requests. From the third request onward, the model receives a stable handle, the original size, and a brief excerpt of the output's head and tail lines, with full pages remaining retrievable.
• Evidence-Preserving Reducer: Build and test logs larger than 4 KiB are routed to a more economical model, GPT-5.6 Luna, which generates a compact receipt. A deterministic verifier ensures the receipt's schema, source hash, exit status, exact quotes, and size are correct. The system reverts to the original log if verification fails, credentials are suspicious, or the receipt offers no size reduction.

Performance and Real-World Implications

SoL-Pi's efficacy was further validated across multiple benchmarks. On Opus 5, the full stack achieved 94.3% of Pi’s score, reducing token traffic by 44.7% and API cost by 33.5%. With GPT-5.6 Sol, it maintained 93.7% of Pi’s score, with a 49.0% token reduction and 33.2% lower cost. These efficiencies translate to estimated hourly savings ranging from $8.75 to $13.50 compared to native Codex and Claude Code harnesses, and $4.36 to $5.71 against the standard Pi agent.

Beyond EdgeBench, SoL-Pi also demonstrated promising results on Terminal-Bench 4, solving 15 out of 63 CPU-only tasks with a 26.3% lower total cost than Pi. In the IMO 2026 Lean 4-verified competition, SoL-Pi matched Pi by passing 3 of 6 problems at the lowest cost per solved problem. While cross-model transfer results are considered preliminary, with mechanisms triggering less often on Opus 5 potentially due to the search being optimized for GPT-5.6 Sol, the overall impact is substantial.

Availability and Future Outlook

SoL-Pi is publicly available on GitHub under the NVlabs umbrella as an MIT-licensed extension, fully compatible with an unmodified Pi release (tested with Pi 0.85.1 and Node.js 22.19 or newer). Its open-source nature facilitates widespread adoption and further development within the AI community. This breakthrough signifies a crucial step toward more sustainable and cost-effective AI-driven software development.