Technology

Nota Inc. Implements 'Full-Stack Optimization' for Agent AI… Reduces Work Time by Up to 47%

Controlling the entire workflow—from models and GPUs to inference, verification, and tool invocation Up to 60% Reduction in Token Output Using the Same Infrastructure Coding Agent Performance Verified… Expanding into Enterprise AI

Lee So-Hyun
2026-11-18 08:50:28
[Edaily Reporter Lee So-Hyun ] #Nota Inc., a company specializing in AI inference efficiency, has unveiled its “full-stack optimization” technology, which goes beyond model lightweighting and GPU optimization to control the workflow of AI agents. While using the same AI model and GPU, the company reduced the processing time of coding agents by up to 47% and output tokens by up to 60%.

Nota Inc.: Comparison of Agent AI Optimization Metrics (Photo: Nota Inc.)


Nota Inc. announced on the 18th that it has developed “full-stack optimization” technology that combines agent execution control with existing AI model lightweighting and GPU kernel optimization.

AI agents repeatedly call models, use tools, and verify results while processing a single request. If unnecessary inference or re-evaluation occurs repeatedly during this process, response times can increase and computational resource consumption can rise.

To mitigate this, Nota Inc. expanded the scope of optimization to include not only the model and GPU but also the agent’s methods for inference, verification, and tool invocation.

Optimizing Even the Agent’s “Harness”

The core of this technology lies in optimizing the “harness” that manages the agent’s workflow. By analyzing recent response lengths, file modification histories, and tool execution results, it selects an inference intensity appropriate for the situation and prevents the repetition of verifications that have already passed or unnecessary re-evaluations.

The model’s response generation process and GPU computation handling have also been improved. By applying a technique that reuses iterative context to “DFlash2-based predictive decoding,” the model’s acceptance rate of predictions was increased, and GPU kernels were tuned to match hardware characteristics. Additionally, by enabling multiple GPUs to share fast execution settings, redundant tuning and latency were reduced.

Processing time reduced by up to 47%… Output tokens reduced by up to 60%

Performance evaluations were conducted using the same AI model and an environment with four NVIDIA RTX PRO 6000 GPUs. The comparison was made against the default configuration of the open-source coding agent “Cline,” using “Terminal-Bench 2.1” and “SWE-bench Multilingual.”

Nota Inc. applied model inference and GPU kernel optimization across the board and measured performance in two modes—“Fast” and “Ultra Fast”—depending on the level of agent control. Fast mode reduced processing time by up to 32% and output tokens by up to 42% while increasing or maintaining the problem-solving rate. Ultra Fast mode reduced processing time by up to 47% and output tokens by up to 60% by minimizing iterative verification and unnecessary re-evaluation.

However, some performance degradation was observed in the process of improving efficiency. In the software bug fixing evaluation, the problem-solving rate in Ultra Fast mode dropped from the baseline of 65.7% to 62.0%. This suggests that differences may arise between processing efficiency and problem-solving rates depending on the task type and optimization level.

Expanding Beyond Coding Agents to Enterprise AI

Nota Inc. plans to expand this technology beyond coding agents to enterprise AI agents and generative AI services in general.

The strategy involves combining agent execution control with existing model lightweighting and hardware optimization to enable customers to process more AI tasks with the same computing resources, and to leverage this into a business model centered on optimized software licenses and technology provision.

Kim Tae-ho, Nota Inc.’s Chief Technology Officer (CTO), emphasized, “As AI agents are increasingly adopted in real-world work, the time and computational cost required to complete a single task become more important than the model’s raw inference speed.” He added, “Nota Inc. is expanding its technological scope beyond model and GPU optimization to include the agent’s inference and verification processes, enabling the processing of more AI tasks on the same infrastructure.”

He added, “Going forward, we will expand our applications to various enterprise AI agents and service environments to reduce customers’ AI operating costs and broaden Nota Inc.’s optimization software business scope.”

Economy

Corporation

IT·Science

Economy

[37th SRE][Issue] BBB-Rated Market Stifled by the JR and JoongAng Crisis

“The bond market should be a universal channel for corporate financing, but right now it’s like an exclusive social club.”This was the point raised by an advisory committee member during a meeting hel…
2026-11-18 10:30:05

Corporation

'Gyochon Makgeolli' Served at APEC Banquet… Available at the Grand Hyatt Hotel

KYOCHON FOOD&BEVERAGE, operator of the Kyochon Chicken franchise, is expanding its sales channels for traditional Korean liquor by introducing its premium makgeolli, “Eunhwasu: A Night of Stargazing 1…
2026-11-18 09:50:52

IT·Science

K-Games Dominate Google Play's 'Game of the Year' List

Google Play announced the “2026 Best of the Year” on the 18th, highlighting the apps and games that were most popular among users over the past year. In South Korea, Nexon’s “Maple Grow” was…
2026-11-18 11:22:31