What's this about?
The authors studied very long agent trajectories and report diminishing returns as extra tokens accumulate. In their comparisons, multiple separate attempts outperformed one prolonged run under a similar budget.
What exactly is a long agent trajectory?
One continuous chain of model actions and observations. Early steps can help, but later steps may revisit errors or overload context instead of making progress.
What's the community saying?
12 Hugging Face daily-paper upvotes.
- Spending more inference budget is not always more effective.
- Fresh attempts and selection can beat a single stubborn run.
- Checkpointing and stopping rules are practical agent-design questions.