Nvidia Open-Sources SoL-Pi to Cut Agent Token Use by Up to 64%

The problem has two layers: immediate waste and long-term difficulty. In real business, AI work now involves hundreds or thousands of agents collaborating over days without human intervention. Long cycles accumulate bloated context and more tool calls. Each model call rereads this history, creating redundancy. AI also lacks human flexibility, sometimes splitting operations that could be completed continuously into two decisions. These scattered inefficiencies burn money. Industry discussions of RSI, or AI improving itself, also carry costs: every attempt consumes tokens, and more than 90% of attempts fail. Blind trial and error can exhaust a company before AI evolves successfully.

Nvidia's approach was to seek efficiency gains while saving money. It did not let AI blindly modify code online. Instead, it had AI propose 152 workflow optimization ideas in batches, then placed them in a strict automated sandbox for selection. The automated screening process is called an automatic research loop. Nvidia also explored three ways for AI to manage AI. The planned-economy model, in which a human or main model fixes every step, was soon abandoned because its rules were too rigid. A centralized-management model used a supervisor AI to write code at runtime, hold meetings, assign tasks to lower-level agents and adjust workflows in real time. It was flexible, but overwhelmed the supervisor and bloated the main model's core codebase. The adopted approach was a shared flash mode: the system maintains only a minimal work template and manual. Each AI experiment copies the template, freely changes code and runs tests, then discards the temporary command code when the experiment ends. This method ran all 152 experiments. Only four effective mechanisms survived and became SoL-Pi's core.

The first mechanism is action fusion. Previously, after changing a line of code, AI would stop to decide whether to run a test, then issue a command to call the terminal, adding an extra inference step. The system now combines fixed consecutive actions such as editing code and running tests into one step, removing that extra decision cost.

The second is online context compression. Transformer architecture means AI memory accumulates linearly, and after each task the model reads earlier context, increasing costs. The optimized system divides large tasks into independent small tasks. After each small task, it compresses that history into a concise version that retains key conclusions. Normally it opens only the summary; when details need verification, it retrieves the original records.

The third is an observation packet. Tool outputs are also fully inserted into context, such as hundreds of lines of test logs. Later discussions of that result must carry the whole passage, consuming space. After optimization, the system stores complete output locally and leaves only a brief summary and index in context. If AI wants details, it retrieves the corresponding section.

The fourth is an evidence-preserving simplifier. When analyzing error logs and troubleshooting, the main model previously had to read from beginning to end. The optimized system uses a cheaper smaller model to distill logs into a diagnosis report, and every conclusion in the report maps back to a specific location in the original text. The main model reads only the condensed report and checks the original if it has questions.

To verify the four mechanisms, Nvidia built 535 verifiable environments. Of these, 495 tasks came from real open-source project bugs found on GitHub. Nvidia restored code repositories to the historical state before the original programmers fixed the bugs and installed dependencies in an offline environment. The correct human fixes were hidden from AI. The other 40 tasks were blind-box challenges: Nvidia wrote a verifier in code and placed AI in an open sandbox with no script, relying on AI to freely explore and trigger the pass condition.

Four harnesses then competed: native framework-driven GPT-5.6 Sol, Claude Opus 5, open-source agent base Pi and Nvidia's SoL-Pi. While maintaining its basic performance, SoL-Pi showed significant savings. Compared with Pi, it reduced token consumption by 45% to 49% and cut development costs by about one-third. Compared with model-native official agent frameworks, its token consumption fell by as much as 64% and API prices by as much as 54%.

Nvidia concluded that efficiency itself is a form of RSI. Rather than only pursuing smarter AI, it argues that making AI work more efficiently first allows saved budget to run more experiments and try more ideas, which in turn accelerates generalization and progress. The paper calls this 'Efficiency for efficiency': efficiency is both a result and a resource for the next round of research. The four mechanisms identify common problems in AI workflows that persist even when models or tasks change. According to the Leiphone report, these ideas emerged from broad exploration, and at this stage breadth matters more than depth. The report added that AI improving AI tools will become more common, and the four optimization ideas can be applied to other AI workflows even without using SoL-Pi.