Post
82
You don't have to run an agent evaluation to the end.
An agent's final outcome is usually visible from its early behavior, so we just stop the run once it's clear.
EarlyEval cuts 13% to 26% of steps and up to 44% of input tokens, and resolve rates move by only 1 to 2 points.
EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction (2609.02783)
https://github.com/inphotoo/earlyeval
An agent's final outcome is usually visible from its early behavior, so we just stop the run once it's clear.
EarlyEval cuts 13% to 26% of steps and up to 44% of input tokens, and resolve rates move by only 1 to 2 points.
EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction (2609.02783)
https://github.com/inphotoo/earlyeval