Nvidia's SoL-Pi rewrites coding-agent harnesses to use up to 49 percent fewer tokens

The four mechanisms the search kept all reduce the text sent to the model. Action Fusion merges two consecutive steps, such as a code edit and the test run after it, into one, removing a whole language model call ^[7]. Anyone who has watched an agent stop to decide whether to run the tests it just wrote will recognise the call being cut. Online Context Compact trims accumulated context after each planning step when it can do so without losing important information ^[8]. ObservationPack archives long tool outputs and sends a short summary on later steps in place of the full text ^[9]. The Evidence-Preserving Reducer gives large error and test logs to a cheaper model to condense, with an automatic check for critical clues that slip through ^[10].

None of these is exotic. The difficulty is coupling. Tool use, context management, verification and abort logic depend on each other, so a change that saves tokens in one step can cause errors in another or push cost into a later phase, according to the researchers as reported by The Decoder ^[16]. That tuning is usually done by people reading long execution traces ^[16].

The evaluation design is the best engineering in the paper. Earlier work found that automatically optimised harnesses overfit to their training tasks and help little on unfamiliar ones ^[6]. The authors walled EdgeBench off from the search. Of its 51 public tasks, 11 were used once to validate finished candidates, and the other 40 were held for final evaluation, with no results fed back ^[5].

The all-four variant is the one to price. It used 49 percent fewer tokens and reached 93.7 percent of Pi's score ^[11]. Divide 0.51 by 0.937 and the variant uses about 0.54 of Pi's tokens per point of score, a cut of roughly 46 percent ^[1]. That figure carries over if you pay per token and partial scores map to partial value on your tasks. Where a failed fix sends the ticket back to an engineer, the 6.3 percent of score given up has to be priced in engineer hours ^[3].

The dollar figures need the same care. The authors estimate savings of $8.75 to $13.50 an hour against the native Codex and Claude Code harnesses, and $4.36 to $5.71 an hour against Pi, at current API prices ^[13]. Pairing low end with low end and high with high, the saving against Pi is 42 to 50 percent of the saving against the native harnesses ^[2]. The remainder comes from starting on Pi at all ^[2]. The Decoder's account does not put a price on the search that found the mechanisms.

Transfer is the open question. Most of the search environments, 495 of them, came from GitHub issue-pull-request pairs ^[4]. Under Opus 5 the mechanisms triggered less often and less aggressively ^[14]. The researchers attribute that to a harness optimised only on GPT-5.6 Sol trajectories ^[14]. The Decoder also reports that results on benchmarks beyond EdgeBench are messier ^[15].

I think the four mechanisms are the part a team can take away. Each can be written by hand and tested against a team's own traces, with a held-out split like the one the authors used ^[5]. The search that produced them ran on GPT-5.6 Sol alone ^[14].