Voices
Agents rot geometrically: every model tested falls from near-perfect to near-zero within sixteen steps
Shubhra Mittal's arXiv 2609.01660 analyzed 10,664 trajectories across nine models from 1.2B to 671B parameters and found task success follows P(n) = p^n on a single per-step reliability parameter. At p=0.95 that gives 59.9% success at 10 steps, 35.8% at 20, and under 1% at 100. Daniel Vaughan turned it into Codex CLI session design rules on September 4, and the counterintuitive result is the one worth acting on: restricting context windows does not alleviate degradation, it steepens it, so cap turns and checkpoint instead of buying more context.
↳ Follow the thread