Stale traces degrade code repair agents despite multiple revisions

The paper isolates trace staleness as a distinct failure mode, separate from agent capability limits. The 22-point penalty from stale execution traces is not about the agent failing to reason better on retry; it's about the agent reasoning from corrupted state. This distinction matters because it's fixable through engineering (state refresh) rather than requiring better models.

This builds directly on 'The Physics of Multi-Turn Long-Horizon Planning' (late July), which showed that explicit world models outperform raw data training for multi-step tasks. That work identified planning as learnable but fragile; this paper reveals one concrete fragility: agents degrade when their internal state representation becomes stale. The finding also echoes the temporal graph generation paper from the same week, which proved that distribution drift causes irreversible degradation in production systems. Here, the drift is temporal rather than distributional, but the core insight is identical: systems deployed on evolving state cannot assume their training assumptions hold across iterations.

If production coding agents (Claude, GPT-4 Code Interpreter, or Devin) adopt explicit trace refresh between revision attempts and report correctness recovery matching the 22-point gap, that confirms the finding generalizes beyond HumanEval. If they don't ship this within six months despite the paper's visibility, it suggests either the penalty doesn't matter at their scale or the engineering cost outweighs the benefit.