Nvidia harness cuts coding-agent tokens nearly in half

On September 26, The Decoder described SoL-Pi, a system from Nvidia researchers. It does not change the model. It changes the harness, the control layer between the model and its environment: how an agent sees state, takes actions, and handles feedback. The report says this is the layer used by systems such as Codex, Claude Code, and OpenClaw. A research agent reads another agent's traces, proposes harness changes, and keeps only those that hold performance about steady while cutting cost.

The search scale in the report is 535 executable environments and 152 directions, including 495 tasks taken from GitHub issue and pull-request pairs and 40 synthetic tests. The process produced more than 3,000 runs and more than 60,000 agent-environment interactions. The researchers also say that more search does not automatically mean a better harness, because automatically optimized harnesses tend to overfit their training tasks.

SoL-Pi separates search feedback from final evaluation. EdgeBench is walled off from search. Of its 51 public tasks, 11 are used once to validate finished candidates. The other 40 are the final evaluation, and those results never feed back into search.

The search produced four mechanisms. Action Fusion merges two consecutive steps, such as a code edit followed by a test, and drops one model call. Online Context Compact runs after each planning step and trims context when it can do so without losing important information. ObservationPack archives long tool outputs and leaves a short summary on later steps. The Evidence-Preserving Reducer sends large error and test logs to a cheaper model and uses an automatic check to catch critical clues that the summary missed.

On EdgeBench's 51 public tasks, The Decoder says SoL-Pi performs about as well as the original Pi harness, with the token reduction depending on the configuration. The leanest variant, using all four mechanisms, uses 49 percent fewer tokens and reaches 93.7 percent of Pi's score. Choosing only the strongest single mechanism beats Pi's score by 5.3 percent and still saves tokens. Across the two variants, token use falls by 44.7 to 49 percent. A figure in the report says the saving versus Codex is 50 percent and versus Claude Code is 54.3 percent, also on EdgeBench. Using API prices at the time, the authors estimate $8.75 to $13.50 an hour versus native Codex and Claude Code, and $4.36 to $5.71 an hour versus Pi. In one test run, total cost fell from $1,339 to $894.

The system was built only on GPT-5.6 Sol trajectories and then applied unchanged to Opus 5. The report says it kept 94.3 percent of Pi's performance with similar savings, but the mechanisms fired less often and less aggressively. The researchers attribute that to optimizing the harness only on GPT-5.6 Sol traces.

The result does not stay "about the same score, about half the tokens" on every benchmark. On 63 CPU tasks from Terminal-Bench 4, SoL-Pi solves 15, while Codex and Pi each solve 18. Total cost was still about a quarter lower than Pi's. On six formally verified Lean 4 tasks from the 2026 International Mathematical Olympiad, the system solved three, at the lowest cost per solved problem. A swarm of 20 SoL-Pi workers cut cost by 26.8 percent versus a comparable Pi swarm on a kernel-optimization test.

Shorter context can reduce reuse of the prompt cache. The authors describe pretraining a harness across many tasks, then using a lean harness to search for the next one, as a vision rather than a result of this study. This article follows The Decoder's account and does not add experimental conditions that the article does not state, because the paper PDF was not the source read here.