1,389 agent trajectories: zero-shot success falls from 82% to 13% as task horizon grows

HaydnBelfield · x · 2026-10-14

Part 6/9 of DeepMind's report thread analyzes time-horizon effects across 1,389 internal Argon agent trajectories: the larger and longer the task, the lower the zero-shot success rate and the more humans intervene, indicating substantial human participation in ML R&D.

The attached chart (Figure 3.16) breaks it down by estimated human time horizon: 82% of tasks under 15 minutes complete without intervention, dropping to 44% for 1-4 hour tasks and just 13% for tasks over a week (83% require human intervention).

Original post →

1,389 agent trajectories: zero-shot success falls from 82% to 13% as task horizon grows

More from coding & agent