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

Nvidia researchers built a system that automatically rewrites a coding agent's harness, cutting token use by 44.7 to 49 percent with little loss on their main benchmark, though results were weaker on unfamiliar tasks.
What it means for founders
- The harness is a cost lever you control. If agent spend is a real line in your budget, this suggests context and tool output handling can matter as much as model choice. An earlier Composio test found cost per solved task varied almost threefold across frameworks running the same model.
- Some of the wins are cheap to copy. Summarizing long tool output, sending logs to a small model and fusing edit and test steps are changes a team can try now without waiting for code.
- Measure solved tasks, not just tokens. The Terminal-Bench drop shows savings can cost you results, so keep an eval suite that catches regressions before a leaner agent ships.
- Watch for a code release and for Codex, Claude Code and other harnesses adopting similar compaction by default, which would change the case for building your own.
The story
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.
Sources
Primary sources
Reporting
More in Research
- OpenAI posts a new batch of math results on GitHub

For founders: Verification is the real cost. A pile of 722 manuscripts is cheap to publish and expensive to check.
The Verge · 42d ago - OpenAI's Navier-Stokes claim awaits Clay review

For founders: Treat lab reasoning claims as unverified until a third party signs off. The most famous AI math claim is still under review a month on.
The Verge · 43d ago - Anthropic says Claude agents found a new CRISPR-like enzyme system in phage DNA

For founders: Agent swarms are becoming a research method. Hundreds of agents screening sequence data in under a day is a template bio startups can study, with token spend…
TechCrunch · 55d ago - OpenAI to work with an independent math panel as it claims over 100 more solved open problems

For founders: Verification is the bottleneck: A lab can now claim more results than outside experts can quickly check.
TechCrunch · 57d ago - Anthropic runs its own wet lab so its AI models can test biology ideas

For founders: Labs are becoming vertical players. A frontier lab with its own bench, a biotech acquisition and pharma partnerships is building capabilities that overlap…
TechCrunch · 60d ago