Nvidia's SoL-Pi Cuts Coding Agent Token Usage by Up to 49% Through Automated Harness Optimization

Nvidia researchers have built a system that automatically optimizes the control layer of coding agents — known as the harness — cutting token usage by nearly half while keeping performance roughly stable, according to a new paper.

The system, called SoL-Pi, targets a cost problem specific to autonomous coding agents: the longer they run unsupervised, the more tokens they burn. Single predictions expand into long chains of tool calls, reasoning steps, and feedback loops. Most existing efficiency work — faster attention kernels, quantization, cheaper model swaps — attacks the cost per token. SoL-Pi instead attacks the harness itself, the layer between the model and its environment used by systems like Codex, Claude Code, or OpenClaw.

How it works

SoL-Pi uses a research agent that watches another agent's execution traces, proposes changes to the harness, and tests candidates in prepared environments. Capability and efficiency checks determine which changes survive. The researchers describe this as a form of recursive self-improvement, though the held-out evaluation set never feeds back into the search process — a safeguard against the overfitting problems seen in earlier automated harness work.

Across 535 executable environments, the system explored 152 candidate directions, generating more than 3,000 runs and over 60,000 agent-environment interactions. The final evaluation used EdgeBench, a benchmark walled off entirely from the search process: 11 of its 51 public tasks were reserved for one-time validation, and the remaining 40 for final scoring only.

Four mechanisms, up to 49% fewer tokens

The search converged on four mechanisms. Action Fusion merges consecutive steps — such as a code edit followed by a test run — into a single model call. Online Context Compact trims accumulated context after each planning step without discarding critical information. ObservationPack archives long tool outputs and substitutes a short summary in later steps instead of resending full text. The Evidence-Preserving Reducer routes large error and test logs through a cheaper model to distill key findings, with an automated verification step catching missed details.

On EdgeBench, the most efficient configuration combining all four mechanisms cut token usage by 49% while retaining 93.7% of the baseline Pi harness's score. A configuration using only the strongest single mechanism beat Pi's score by 5.3% while still saving tokens. Across configurations, token savings ranged from 44.7% to 49%. SoL-Pi saved 50% versus native Codex and 54.3% versus Claude Code on the same benchmark, according to Nvidia.

In dollar terms, the authors estimate savings of $8.75 to $13.50 per hour compared to native Codex and Claude Code harnesses, and $4.36 to $5.71 per hour compared to the baseline Pi harness, based on current API pricing. One test run's total cost dropped from $1,339 to $894.

The harness was built using GPT-5.6 Sol and then applied unmodified to Opus 5, retaining 94.3% of Pi's performance with similar token savings — though the mechanisms triggered less aggressively, which the researchers attribute to the harness being tuned only on GPT-5.6 Sol trajectories.

Mixed results outside EdgeBench

Results on other benchmarks were less clean. On 63 CPU tasks from Terminal-Bench 4, SoL-Pi solved only 15 tasks versus 18 each for Codex and Pi, though total costs were about 25% lower than Pi's. On Lean 4 tasks from the 2026 Math Olympiad, SoL-Pi solved three of six problems at the lowest cost per solved problem. In a kernel-optimization test using a 20-worker swarm, SoL-Pi cut costs by 26.8% compared to an equivalent Pi swarm while producing the best result.

The researchers note a trade-off: shorter context can reduce prompt cache reuse, a cost factor that matters given that, according to OpenRouter analyst Peter Walker, agentic token usage has grown 14x since February 2026, with nearly 70% of that volume coming from cached prompts.

What this means

SoL-Pi is a research system, not a product, and its own numbers show the gains are uneven outside the benchmark it was tuned against — EdgeBench performance improved or held steady, but Terminal-Bench 4 task-solving actually declined even as costs fell. That's consistent with the field's broader difficulty: automated harness optimization tends to overfit to its training distribution, and Nvidia's strict separation of search and evaluation data is itself an acknowledgment of that risk.

The more durable signal here is where the cost lives. Composio's August test showed cost-per-task varying by nearly 3x across agent frameworks running the identical model, meaning harness engineering — not model choice — is already a bigger lever on agent economics than many teams assume. If Nvidia's approach or something like it gets productized, the next round of agent cost wars may be fought over control logic, not weights.