NVIDIA's SoL-Pi halves agent token traffic without touching the model
SoL-Pi is a set of four mechanisms bolted onto the Pi harness that cut recorded token traffic by 44.7 to 49.0 percent while holding performance level with plain Pi across GPT-5.6 Sol and Opus 5. API costs drop by about a third. The authors put the savings at $8.75 to $13.50 an hour versus native Codex and Claude Code. No retraining, no fine-tune, no new model. Fourteen authors out of NVIDIA, arXiv 2609.20519, submitted September 17, MIT licensed at https://github.com/NVlabs/SoL-Pi.
The four mechanisms are worth naming because each one attacks a different kind of waste. Action Fusion lets an edit or write run its follow-up validation command inside the same tool call, killing a whole class of round trips. ObservationPack turns repeated large text results into stable handles you can page back into with exact recall, so the same file dumped three times costs you once. The Evidence-Preserving Reducer compacts long diagnostic logs into receipts, but only admits a compaction if every retained quotation still matches the archived source, which is the rare version of summarization that cannot silently invent. Online Context Compact lets completed plan steps become eligible for compaction after economic and window-pressure checks rather than on a fixed trigger.
That third mechanism is the interesting one and it is doing more than saving tokens. The standing problem with agent compaction is that the agent writes a summary and then reads it back as trusted context, with nothing checking that the summary reflects what actually happened. Requiring every retained quote to match the archive is a verification step in a place that normally has none. It is a small idea that generalizes far past cost.
The evaluation is on EdgeBench, 51 tasks, and the claim the authors care about most is that the improvements transfer beyond the setting they were developed in. Treat the 44.7 to 49.0 percent as a number from a lab that ships harnesses and would like you to believe harnesses matter, which is fine, because the mechanisms are legible enough to reimplement and check yourself. What lands is the direction: the cheapest capability gains available right now are not in the weights.
Related: the 176-config harness study https://clauday.com/article/5e057138-f455-4504-b2e9-e6cbcbc1c0d6
← Back to all articles
The four mechanisms are worth naming because each one attacks a different kind of waste. Action Fusion lets an edit or write run its follow-up validation command inside the same tool call, killing a whole class of round trips. ObservationPack turns repeated large text results into stable handles you can page back into with exact recall, so the same file dumped three times costs you once. The Evidence-Preserving Reducer compacts long diagnostic logs into receipts, but only admits a compaction if every retained quotation still matches the archived source, which is the rare version of summarization that cannot silently invent. Online Context Compact lets completed plan steps become eligible for compaction after economic and window-pressure checks rather than on a fixed trigger.
That third mechanism is the interesting one and it is doing more than saving tokens. The standing problem with agent compaction is that the agent writes a summary and then reads it back as trusted context, with nothing checking that the summary reflects what actually happened. Requiring every retained quote to match the archive is a verification step in a place that normally has none. It is a small idea that generalizes far past cost.
The evaluation is on EdgeBench, 51 tasks, and the claim the authors care about most is that the improvements transfer beyond the setting they were developed in. Treat the 44.7 to 49.0 percent as a number from a lab that ships harnesses and would like you to believe harnesses matter, which is fine, because the mechanisms are legible enough to reimplement and check yourself. What lands is the direction: the cheapest capability gains available right now are not in the weights.
Related: the 176-config harness study https://clauday.com/article/5e057138-f455-4504-b2e9-e6cbcbc1c0d6
Comments