NVIDIA-Led Team Releases SoL-Pi, Cutting Coding Agent Token Traffic by Up to 49%

The extension ships on GitHub under the NVlabs organization under an MIT license and runs on an unmodified Pi release. It was tested with Pi 0.85.1 and Node.js 22.19 or newer.

Most efficiency work lowers the cost per token through faster kernels, quantization or cheaper models. SoL-Pi instead reduces how many tokens a task consumes. The harness is the layer that handles tool calls, context, observations and delegation. Tuning a harness by hand is slow and its parts are coupled, so a fix in one place can push cost into later steps. Meta-Harness and similar systems automate that work, but a recent study found that evolved harnesses can overfit their search tasks and give only marginal gains on unseen ones, according to the report.

In SoL-Pi's search, a research AI observes execution traces from a separate agent running base Pi, then proposes harness changes and tests them. The search covered 152 proposed directions across six families, namely context, progress, tools, delegation, prompt and policy, and improvement and evaluation, plus 535 executable environments, of which 495 were built from GitHub issue-pull request pairs and 40 were synthetic tasks with executable verifiers. It ran more than 3,000 runs and over 60,000 agent-environment interactions. Each search is a disposable, isolated loop that follows the autoresearch cycle, extended with a Ralph Loop implementation step and an independent reviewer. Acceptance rules are fixed before a search starts and the optimizer cannot change them: every capability metric must stay within a predeclared tolerance, and a candidate must also improve at least one efficiency metric. EdgeBench stays held out, with 11 of its 51 public tasks used for one-way acceptance of frozen candidates and 40 for final evaluation, and held-out results never feed back into the search.

The four surviving mechanisms work on different parts of a run. Action Fusion merges a file edit and a separate test, build or run command into one tool request and returns both outcomes in one observation, removing a model round trip. Online Context Compact tracks plan steps through update_plan; when a step completes, the harness estimates how many requests remain and compares projected input savings with the extra cost of rewriting the prompt cache, invoking Pi's native compaction when that gate passes or when context nears the window limit. ObservationPack archives tool outputs above 10 KiB locally 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 excerpt of head and tail lines, and exact pages stay retrievable through the handle. Evidence-Preserving Reducer sends build and test logs of at least 4 KiB to a cheaper model, GPT-5.6 Luna at high, which writes a compact receipt; a deterministic verifier checks the receipt's schema, source hash, exit status, exact quotes and size, and the harness falls back to the original log when verification fails, credentials are suspected, or the receipt is not smaller.

The full stack was built on GPT-5.6 Sol and moved to Opus 5 without further search. On Opus 5 it keeps 94.3% of Pi's score while cutting token traffic by 44.7% and API cost by 33.5%; on GPT-5.6 Sol it keeps 93.7% of Pi's score with 49.0% fewer tokens and 33.2% lower cost. A performance configuration uses the best single mechanism for each backend, ObservationPack on GPT-5.6 Sol and Action Fusion on Opus 5, and lifts scores 5.3% and 12.8% above Pi. On GPT-5.6 Sol the full stack raises cache-write traffic from 0.0141 billion to 0.0316 billion tokens, while total cost still falls from $1,339 to $894. The paper estimates hourly savings of $8.75 to $13.50 against the native Codex and Claude Code harnesses, and $4.36 to $5.71 against Pi.

Outside EdgeBench, SoL-Pi solves 15 of the 63 CPU-only tasks in Terminal-Bench 4, against 18 for both Codex and Pi, while lowering total cost by 26.3% versus Pi, at $211.12 against $286.45. On IMO 2026 problems verified in Lean 4, it passes 3 of 6, matching Pi and reaching the lowest cost per passed problem at $20.90, while Codex passes 5.