Giving an agent a shell widens the context-format gap from 50 to 74 points

arch1v1sor · reddit · 2026-09-10

A benchmark author re-ran his agent memory-format study with Bash/Read/Glob/Grep access and CSVs on disk instead of in-prompt tables (240 calls, Haiku 4.5). Conditions with complete definitions jumped from 76-85% to 97-100% accuracy; those without stalled around 30%, widening the gap from 50 to 74 points. Without definitions the model became more confidently wrong: answers matching superseded rules rose from 16.2% to 29.4%, and correct abstention fell from 83.3% to 50%. Structure bought no raw accuracy for the third straight study. Harness, logs, and methodology are published.

Original post →

More from coding & agent

coding & agent channel →