SoL-Pi: Scaling Auto-Research Loops for Efficient Agent Harnesses — NVlabs Open-Source Optimization Framework 2026
Introduction You’re watching your Claude Code session burn through tokens. The agent writes a file, runs validation, writes another file, runs validation again. Each edit is followed by the same pytest command you’ve seen it run twelve times already. The context window fills with repeated outputs. By hour three, your API bill is climbing and you’re wondering if the agent could have finished the work faster if it hadn’t been replaying the same observations. ...