An agent’s plumbing, not its weights, is where four new efficiencies come from. SoL-Pi targets the harness layer of the open-source Pi coding agent, the code that handles tool calls, context, observations and delegation, and a research AI running auto-research loops proposed its mechanisms rather than a human.
The team spans NVIDIA, NTU and MIT. On EdgeBench, a 51-task evaluation, token traffic drops 44.7 to 49 percent against base Pi and API cost falls roughly 33 percent, with scores holding close on GPT-5.6 Sol and Opus 5. The MIT-licensed extension sits on GitHub under NVlabs, installs on an unmodified Pi release, and is tested with Pi 0.85.1 on Node.js 22.19 or newer.
Breadth was the point. The optimizer floated 152 directions across six families covering context, progress, tools, delegation, prompt and policy, and improvement, then tested them across 535 executable environments, 495 assembled from GitHub issue and pull request pairs plus 40 synthetic tasks with verifiers, running more than 3,000 trials and 60,000 agent-environment interactions. Acceptance rules were frozen before the search, and EdgeBench stayed out of it.
Three mechanisms produce most of the savings. Action Fusion collapses a file edit and its follow-up test into a single tool request with one combined observation. Online Context Compact decides when rewriting the prompt cache is worth the cost. ObservationPack stores tool output larger than 10 KiB locally and gives the model a stable handle instead.
