Nvidia cuts code-agent token traffic in half with SoL-Pi

A research team from Nvidia, NTU and MIT has released SoL-Pi, an MIT-licensed extension to the open-source code agent Pi that reduces token traffic by 44.7% to 49% and API cost by 33%, while preserving 93% to 94% of the original performance. The mechanisms were discovered by an AI running automated research loops in the harness layer, not through manual optimization.

Why the harness, not the model

Most efficiency work targets cost per token — faster kernels, quantization, cheaper models. SoL-Pi moves in the opposite direction: it shrinks the number of tokens a task consumes in the first place. The harness is the layer that manages tool calls, context, observations and delegation. Hand-tuning it is slow and its components are coupled; a fix in one place can push cost into later stages. Systems such as Meta-Harness have already automated this search, but recent research shows they tend to overfit on the search tasks themselves and deliver only marginal gains on unseen tasks.

How the search worked in practice

An AI researcher observed execution traces of a base Pi agent, proposed harness changes and tested them. The search covered 152 proposed directions across 6 families — context, progress, tools, delegation, prompt & policy, improvement & evaluation — across 535 execution environments, 495 built from GitHub issue-PR pairs and 40 synthetic tasks with executable validators, for more than 3,000 runs and 60,000-plus agent-environment interactions. Each search is an isolated, single-shot loop that follows an extended auto-research cycle with a Ralph Loop implementation phase and an independent critic. Acceptance rules were fixed in advance; the optimizer cannot change them. Every capability metric must remain within a declared tolerance band, and the candidate must improve at least one efficiency metric. EdgeBench was held out: 11 tasks for one-shot acceptance of frozen candidates, 40 for final evaluation. Held-out results do not feed back into the search.

The four mechanisms that survived screening

Action Fusion merges a file edit and a test-or-build run into a single tool request and returns both results in one observation, saving a full model round-trip. Online Context Compact tracks plan steps via update_plan; when a step ends, the harness estimates how many requests remain, compares the expected input savings against the additional cost of rewriting the prompt cache, and triggers Pi's native compression when the gate passes or the context nears the window limit. ObservationPack locally archives tool outputs larger than 10 KiB and sends them in full for the next two provider requests; from the third request onward the model sees a stable handle, the original size and a short head-and-tail snippet, while exact pages remain accessible through the handle. Evidence-Preserving Reducer sends build and test logs of 4 KiB or more to a cheaper model, GPT-5.6 Luna in high mode, which writes a compressed receipt; a deterministic validator checks schema, source hash, exit status, exact citations and size. The harness falls back to the original log in three cases: when validation fails, when credential leakage is suspected, or when the receipt is not smaller.

Results on EdgeBench and next steps

The full stack was built on GPT-5.6 Sol and ported to Opus 5 without additional search. On Opus 5 it retains 94.3% of Pi's score with 44.7% fewer tokens and 33.5% less cost; on GPT-5.6 Sol it retains 93.7% with 49% fewer tokens and 33.2% less cost. The code is available on GitHub under NVlabs, runs on Pi 0.85.1 without modifications and requires Node.js 22.19 or later.