Stanford Meta-Harness Hits 76.4% on TerminalBench 2.0 — Environment Bootstrapping Saves 2-5 Agent Turns Per Task

Stanford IRIS Lab / GitHub·high signal

Stanford IRIS Lab's Meta-Harness extends the KRAFTON Terminus-KIRA framework with environment bootstrapping: before the agent loop starts, it snapshots the sandbox (working directory, available languages, package managers, memory) and injects it into the initial prompt. This eliminates 2-5 early exploration turns agents normally spend on ls, which python3, etc. Running Claude Opus 4.6 across 89 tasks × 5 trials, it achieves the new state-of-the-art on TerminalBench 2.0.