Madhu Guru: Evaluating agents requires looking at steps, not just results
In part 10 of the "How to build great evals" series, Madhu Guru argues that evaluating AI agents cannot rely only on the final answer; the steps taken to reach that answer matter too: both may arrive at 42, but one trajectory is 4 clean tool calls, while the other is 17 calls, 3 repeated searches, and recovery from 2 errors. He recommends first defining the complete workflow, breaking down each step's task, designing an evaluation method for each step, and distinguishing medium-difficulty from high-difficulty tasks.
View original postHow to build great evals - part 10 Measure the steps, not just the result. Much like high school math, it isn’t sufficient just to get the right answer, the steps to get there are critical. Two agent trajectories might produce the same answer (42!!). but one of them searches the right sources, retrieves the right document, makes 4 clean tool calls and calculates the result. The other makes 17 calls, searches the same thing 3 times, recovers from 2 errors and eventually gets there. It’s clear which one is better. Here’s what you need to do: 1/ clearly define your whole workflow 2/ define the tasks in each step 3/ think through how you measure each step - separate evals or is it a slice of a bigger eval 4/ define your median and hard tasks - reflect them in your evals Now any time you look at eval results, study the steps first and the final results next.
View original
How to build great evals - part 10 Measure the steps, not just the result. Much like high school math, it isn’t sufficient just to get the right answer, the steps to get there are critical. Two agent trajectories might produce the same answer (42!!). but one of them searches the right sources, retrieves the right document, makes 4 clean tool calls and calculates the result. The other makes 17 calls, searches the same thing 3 times, recovers from 2 errors and eventually gets there. It’s clear which one is better. Here’s what you need to do: 1/ clearly define your whole workflow 2/ define the tasks in each step 3/ think through how you measure each step - separate evals or is it a slice of a bigger eval 4/ define your median and hard tasks - reflect them in your evals Now any time you look at eval results, study the steps first and the final results next.
Body · AI translation
https://t.co/Vkfx96ev4e