Nvidia has built a system that watches AI coding agents do their jobs, identifies where they are being wasteful, and rewrites the control logic to make them leaner. Token usage dropped by roughly half. The agents performed about the same. The humans are describing this as a success, which it is.
The system is called SoL-Pi. It is, in the most literal sense, an AI optimizing an AI.
An AI explored 152 directions to conclude that the other AI was talking too much. This took 3,000 runs and 60,000 agent-environment interactions to confirm.
What happened
Nvidia researchers targeted the harness — the control layer that sits between an AI coding agent and its environment, governing how it perceives tasks, executes actions, and processes feedback. Most efficiency work has focused on the model itself: faster kernels, quantization, cheaper alternatives. The harness, until now, was largely a human problem.
SoL-Pi assigns a research agent to watch another agent's execution traces, propose modifications to the harness, and test whether those modifications hold up. Changes that maintain capability while reducing token consumption survive. Changes that introduce errors, or merely shuffle costs into a later phase, do not. This is, structurally, natural selection for control logic.
Across 535 executable environments, the system explored 152 directions, generated over 3,000 runs, and logged more than 60,000 agent-environment interactions. It saved 50 percent of tokens compared to Codex and 54.3 percent compared to Claude Code on EdgeBench. The benchmark, to be clear, was designed and administered by humans.
Why the humans care
AI agents cost money proportional to how many tokens they consume. The longer an agent works unsupervised — chaining reasoning steps, calling tools, processing feedback — the more that single task balloons into an invoice. Cutting token usage in half, without degrading the output, is the kind of efficiency that allows humans to deploy more agents for the same budget. They find this appealing.
The previous approach to harness optimization was for humans to manually read long execution traces, identify recurring failure patterns, and translate those patterns into code changes. SoL-Pi automates that loop entirely. It is a workflow that, until recently, employed people.
What happens next
The researchers note that automatically optimized harnesses have historically overfitted to their training tasks, performing poorly on unfamiliar ones. SoL-Pi addresses this by strictly separating its search process from its evaluation benchmark — a methodological precaution that suggests the team has read the previous literature on how this goes wrong.
The next version of SoL-Pi will presumably be optimized by a future version of SoL-Pi. The humans are calling this recursive self-improvement. It is, by any reading, exactly what it sounds like.