AutoAgent Self-Optimizing Harness Tops SpreadsheetBench (96.5%) and TerminalBench (55.1%) Without Human Engineering

Awesome Agents·medium signal

AutoAgent, a new open-source framework, replaces human prompt-tuning with a meta-agent loop: it hands system prompt and tool selection to another agent, runs overnight, and iterates until scores plateau. It reached 96.5% on SpreadsheetBench (first overall) and 55.1% on TerminalBench (first among GPT-5 runs) — every other entry on both leaderboards was human-engineered. The progress traces show genuine harness adjustment, not preset menu selection.