Coding Agent Defaults Need Benchmark Ledgers

Our Python data pipelines for WaterDoctor, one of our deep-tech investees, flew with a default agent — clean, self-contained Pandas edits and a high solve rate in a single zero-shot pass on tasks that at least resembled HumanEval’s short-function setup, with no retrieval and no human edits to the output. Same isolation pattern as the benchmark, different workload.