NVIDIA open-sources an agent harness that redesigned itself
NVIDIA's research arm published the code for a loop where agents studied other agents' transcripts and cut their own token bill. Meanwhile South Korea's central bank measured who actually gets the AI jobs.
NVLabs released SoL-Pi, an MIT-licensed extension for the Pi coding agent that packages four efficiency mechanisms discovered by automated research loops rather than by engineers. It installs on an unmodified Pi release, and every mechanism is opt-in and off by default. The four: Action Fusion bundles a file edit with the validation command that always follows it into one local sequence, deleting the model turn in between; ObservationPack archives oversized tool output to local disk and leaves a handle plus a short excerpt in context, paging the detail back only when a later turn needs it; Evidence-Preserving Reducer hands a thousands-of-lines log to a cheap model first and accepts its summary only after checking that every quoted line appears verbatim in the archived original; Online Context Compact treats a finished subtask as a compaction candidate and fires only when projected future savings beat the cost of rewriting the context, which breaks KV-cache reuse. NVIDIA reports 45% to 49% fewer tokens and roughly a third lower cost versus stock Pi while keeping about 94% of its average score, and 35% to 64% fewer tokens with 50% to 54% lower billed cost against stock Codex and Claude Code harnesses — Chinese coverage pegs the practical saving at $8.75 to $13.50 an hour on long research runs. The mechanisms are all the same trick applied at different points in the loop: a large observation that enters context once gets re-billed every turn afterward, which is why the context window is the real cost driver, not the model.
What makes the release interesting is the method behind it. NVIDIA frames harness improvement as an open-ended recursive self-improvement search: 152 proposed directions, screened before rollout budget was spent, run through independent loops against 535 executable environments — 495 mined from GitHub issue-to-PR pairs with the fix reverted and the maintainer's regression test withheld, plus 40 synthetic verifier-defined tasks. Candidates had to clear two gates: every capability metric within a predeclared tolerance, and at least one efficiency metric improved. About 40 ideas produced one surviving mechanism. Held-out validation ran on EdgeBench, chosen because it is one of the few public benchmarks that sustains two- to twelve-hour trajectories, long enough for per-turn waste to accumulate measurably. Three honest caveats: every number is NVIDIA's own, no independent reproduction exists yet, the arXiv preprint is only "coming soon," and the team's own project page concedes the multi-agent swarm comparison is one nonrandomized trial per condition. The stated endgame — a cheaper harness making the next round of auto-research cheaper, which they call "efficiency for efficiency" — is explicitly a research vision, not a result.
The Bank of Korea reported that four-fifths of the country's new AI-exposed jobs landed in the capital region. In the three years since generative AI took off, 247,000 jobs with high exposure to it appeared, and 80.8% of them — 199,000 — were in Seoul, Incheon, Gyeonggi and Sejong. Of the 105,000 new jobs exposed to agentic AI in 2025, 88.3% were in the capital area. Physical AI exposure runs the other way, concentrated in Gyeongsangbuk-do, Jeonnam, Chungbuk and Ulsan, which is the risk the report flags for the provinces. The mechanism is plain occupational arithmetic: office, sales, managerial and professional roles are all denser in the capital region, and so is the infrastructure — 71.9% of knowledge-service workers and 74.4% of semiconductor manufacturing already sit there. One detail cuts across the geography: employment in generative- and agentic-AI-exposed roles fell for workers in their 20s in both regions, which matches what the Korean payroll data showed this summer — the entry rung goes first regardless of postcode. Jung Min-soo, who led the research, said physical AI's spread will take time because of missing physical data, immature capability and user trust.
What to watch: whether SoL-Pi's four patterns get ported into Claude Code and Codex by third parties before NVIDIA says anything about supporting them, and whether the Bank of Korea's regional AI hub recommendations survive into Korea's next budget.
If the cheapest way to cut your agent bill is to stop re-sending the same file every turn, how much of your current AI spend is transfer waste rather than thinking? Tell us in the comments.
Sources: NVIDIA NVlabs SoL-Pi (GitHub)
