NVIDIA's SoL-Pi lets coding agents research their own efficiency
NVIDIA, MIT, and NTU's SoL-Pi has a coding agent generate 152 efficiency research proposals and keep four, cutting tokens nearly in half on the authors' EdgeBench suite.
NVIDIA, MIT, and NTU researchers have a new answer to the ballooning cost of agentic coding: let the agent research the problem itself. A paper published today, SoL-Pi, describes a framework in which a coding agent autonomously generates and tests its own efficiency improvements. The results are striking: on the authors' 51-task EdgeBench suite, the self-improved agent used 44.7 to 49 percent fewer tokens than the baseline Pi agent while keeping 93.7 percent of its score, at roughly a third lower cost.
The team, which includes MIT's Song Han, started by having the agent produce 152 research proposals for making agentic coding cheaper. Four mechanisms survived testing: Action Fusion, which batches related tool calls; Online Context Compact, which shrinks context as a session grows; ObservationPack, which compresses tool outputs; and Evidence-Preserving Reducer, which trims reasoning traces without losing the evidence the agent needs to stay correct.
The framing matters. Most agent-efficiency work is human researchers hand-designing tricks like prompt caching or context compaction. SoL-Pi turns the agent into the researcher: propose, implement, measure, keep what works. The authors released the code under the MIT license at NVlabs/SoL-Pi, with an interactive research blog at nvlabs.github.io/SoL-Pi, and the system is tested against Pi 0.85.1 on Node 22.19 and up.
A few caveats are worth keeping in mind. EdgeBench is the authors' own benchmark, so the 93.7-percent score retention is a vendor claim until independent teams reproduce it. Token reduction also does not map one-to-one to bill reduction, since providers price input, output, and cache reads differently. Still, the direction is right: as agents run longer sessions, the cost of context dwarfs the cost of the model call, and anything that systematically shrinks context is worth watching.
The timing is notable too. This lands the same week Anthropic cut Opus pricing 20 percent and OpenAI shipped cheaper models, which suggests the whole industry has decided the next frontier is not just smarter models but cheaper runs. SoL-Pi is the research-flavored version of that bet, and it is unusually practical: the code is out, the license is permissive, and the mechanisms are small enough to port into other harnesses.
One design choice deserves credit: the four mechanisms are composable, so an agent can stack all of them rather than picking one. That is how the savings add up to nearly half of all tokens. The honest accounting question is what the research loop itself costs. Generating 152 proposals and testing them burns a lot of tokens up front, which only pays off if the winning mechanisms get reused across many sessions. For a lab or a company running agents all day, that math works. For a single weekend project, it would not, and the paper does not claim otherwise.