Nvidia's SoL-Pi cuts coding agent token use by nearly half

Nvidia researchers have built a system that automatically rewrites the harness of a coding agent, the layer that decides what the model sees, which tools it runs and how it handles feedback, and they report that it cuts token use by 44.7 to 49 percent with little loss in results. The paper calls the system SoL-Pi, after the open Pi harness it starts from.

How it works

Most work on agent costs targets the price of each token through faster serving, compressed models or cheaper models. SoL-Pi goes after the number of tokens instead. A research agent reads another agent's execution traces, proposes changes to the harness and tests each one in prepared environments, keeping only changes that pass both capability and efficiency checks. To avoid tuning to the test, the benchmark used for the final scores, EdgeBench, was kept out of the search.

The search settled on four changes:

• merging steps that usually run back to back, such as an edit followed by a test, into a single model call;
• trimming the accumulated context after each planning step when nothing important would be lost;
• storing long tool outputs and passing only a short summary on later turns;
• handing bulky error and test logs to a smaller, cheaper model to condense, with a check that key clues survive.

The results

With all four switched on, the agent used 49 percent fewer tokens and reached 93.7 percent of the original harness's score. Using only the strongest single change scored 5.3 percent above the original and still used fewer tokens. At current API prices, the authors put the saving at $8.75 to $13.50 for every hour of agent work compared with the stock Codex and Claude Code harnesses. The harness was built entirely from GPT-5.6 Sol runs yet kept 94.3 percent of performance on Opus 5 without changes, although its mechanisms fired less often there.

Away from its home benchmark the results were weaker. On 63 Terminal-Bench 4 tasks it solved 15, against 18 for both Codex and Pi, though its total cost was about a quarter below Pi's. Shorter context can also reduce prompt cache reuse, and separate research has found that compressing context tends to drop user instructions.

What we don't know yet

It is not yet clear whether the savings hold on everyday production codebases rather than benchmark tasks, or whether the harness will be released for others to use. The authors present a harness that repeatedly improves its own successor as a direction for future work, not a result of this study.