Removing the Thought Field Hurt My AI Agent Evaluation

0:00This one prompting trick increased my agent accuracy from 84 to nearly 90

0:05with no change in cost For context I'm building a fully autonomous AI

0:11trading agent The way the agent works is you type in a plain English

0:15sentence then the agent does a few things Number one it creates a plan

0:20to accomplish whatever goal is associated with that sentence So you're trying to create

0:25an algorithmic trading strategy that trades Nvidia options It's going to create a plan

0:30to test out different trading strategies Step two is we execute the plan using

0:35the agent executor And what this looks like is an LLM prompt that responds

0:39with two things a thought of the current state where we are in the

0:43process in the next best possible action I wanted to see if I could

0:47save some money or time removing the thought field and only responding with the

0:53action What I found actually shocked me Removing the thought field decreased my agent's

0:58performance from 89 9 to 85 7 I saw no real measurable change

1:04in the costs but it was a little bit that I did save some

1:09time and latency

1:10This makes sense because I'm simply outputting less tokens but these little bit of

1:14tokens don't have a real impact on my costs If you want to understand

1:17how performance evaluation then like and follow for part two