NVIDIA Researchers Released New AI Agent Harness

The SoL-Pi software reduces token traffic and API costs while maintaining high performance for AI models.

Updated on Sept. 19, 2026 in Artificial Intelligence

Isometric editorial illustration of a segmented steel conduit acting as a junction, representing streamlined data flow in automated AI research. — NVIDIA researchers released SoL-Pi, a new agent harness designed to streamline data traffic and reduce API costs for autonomous AI systems. AI Illustration. Upload story photo >

Live Poll

Would you trust an automated AI research agent to complete tasks for your business?

NVIDIA researchers have unveiled SoL-Pi, a new agent harness designed to regulate information flow between language models and their environments. The system utilizes four specific mechanisms to significantly improve efficiency in automated AI workflows.

Why it matters

By reducing token traffic and API costs, this research addresses a major barrier to the scaling of complex autonomous AI agent systems. The development offers a more sustainable pathway for conducting high-volume, automated research tasks.

SoL-Pi achieved 93.7% of baseline performance while reducing token usage by up to 49%. The system, which is released under an MIT license, successfully solved 15 tasks on Terminal-Bench 4 and passed half of the problems on the IMO 2026 test.

The players

NVIDIA

NVIDIA is a global technology company that specializes in the design of graphics processing units and the development of high-performance artificial intelligence infrastructure.

The details

The harness manages workflows through four integrated mechanisms, including an Action Fusion tool that combines edits and tests into a single call. This streamlined communication between the model and the environment allows the system to maintain high accuracy despite the significant reduction in API data throughput.

September 17, 2026: The NVIDIA research team published the SoL-Pi agent harness.

The Tech Race

This development represents a shift toward optimizing agentic workflows, moving the industry beyond raw model size toward operational efficiency. It follows a pattern set by Terminal-Bench 4, which prioritizes real-world tool use capabilities over traditional text-only benchmarks.

Users and developers can expect lower costs and higher efficiency when implementing complex AI agent tasks. The MIT-licensed software allows developers to integrate these efficiency-boosting workflows directly into their own technical applications.

The takeaway

This research highlights that smarter workflow management can be as impactful as increasing model parameters for AI performance. Developers should monitor open-source releases for similar tools to optimize their own AI-driven resource consumption.