Nvidia introduced the SoL-Pi system to reduce tokens for AI agents

Researchers at Nvidia have developed the SoL-Pi system, which reduces the token usage of coding AI agents by nearly 49%. At the same time, the performance of the models remains at the same level due to the optimization of the control layer between the neural network itself and its working environment.

To create the system, a specialized research agent tested 152 approaches in over 3,000 runs. The optimization affects the so-called harness - the environment and set of instructions through which the agent interacts with the file system and terminal.

In key benchmarks, the reduction in token usage reached the claimed 49 percent, but in other tests, the indicators were more modest. The development is designed to significantly reduce the financial costs of performing long tasks of writing and debugging code by autonomous agents.

The new approach allows for more efficient use of the context window and reduces overhead costs for service requests, which is critical for scaling AI-based software development systems.