Nvidia researchers unveil SoL-Pi, a system that automatically optimizes coding-agent harnesses to cut token usage by up to 49% with minimal performance loss
Nvidia researchers published a paper describing SoL-Pi, a system where a research AI agent analyzes coding-agent execution traces, proposes changes to the agent "harness" (the control layer managing tool use, context, verification and abort logic), and keeps only changes that preserve performance while reducing token/cost usage. Tested across 535 executable environments (152 explored directions, 3,000+ runs, 60,000+ agent-environment interactions), it produced four mechanisms (Action Fusion, Online Context Compact, ObservationPack, Evidence-Preserving Reducer). On EdgeBench, the most efficient configuration cut token usage by 49% while retaining 93.7% of the original Pi harness's score; a performance-focused variant beat Pi's score by 5.3% while still saving tokens. Results were more mixed on other benchmarks (Terminal-Bench 4, IMO 2026 Lean 4 tasks, kernel optimization swarms).
Entities: Nvidia, SoL-Pi, Codex, Claude Code, OpenClaw, EdgeBench
0 primary
What happened
Nvidia researchers published a paper on SoL-Pi, a system that uses a research AI agent to automatically rewrite the "harness" (the control layer for tool use, context and verification) around coding agents, keeping only changes that cut token usage without hurting performance. Across 535 test environments, its best efficiency configuration cut token usage by 49% on EdgeBench while retaining 93.7% of the baseline harness's score; a performance-tuned variant beat the baseline score by 5.3% while still saving tokens. Results were weaker and more mixed on other benchmarks, including Terminal-Bench 4, IMO 2026 Lean 4 tasks and kernel optimization swarms.
Why it matters
This is a research paper, not a released product, so nobody can adopt SoL-Pi today. If the approach holds up, it points to a second lever for cutting AI agent costs beyond swapping to cheaper models: optimizing the harness itself, which Nvidia estimates could save several dollars per hour versus running Codex, Claude Code or a baseline harness at current API prices. That would matter most to enterprises and developers running agents unsupervised for long stretches, where token costs compound.
What is noise
The 49%/93.7% headline number comes from the single best-case benchmark (EdgeBench); performance was mixed or weaker on Terminal-Bench 4, Lean 4 proof tasks and kernel optimization, which the coverage underplays. Auto-optimized harnesses have a known failure mode of overfitting to their test environments, and there's no released tool, code or direct paper link in this coverage to verify the claims independently; the source (The Decoder) is secondary, not primary reporting.
Watch next
- 01Whether Nvidia or the authors release the SoL-Pi code, harness, or paper publicly for independent verification
- 02Third-party replication or benchmark results outside Nvidia's own 535-environment test set, especially on Terminal-Bench-style tasks
- 03Whether any coding-agent vendor (Anthropic, OpenAI, Composio) adopts or cites harness-level optimization techniques in a shipped product within the next 2-3 months
Coverage
1 storyMore capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680
- OpenAI launches GPT-6 Sol and Luna at half the token price of GPT-5.6, with roughly flat intelligence scores per independent analysis22 Sept 202680
- Anthropic threat report: Claude abused for malware, drone/missile software, mass surveillance, and industrial-scale distillation by Chinese AI labs11 Sept 202680