I am wondering about the optimal thinking budget when building an agent that runs on gemini 2.5 pro. In my case its a simple ReAct agent with only a couple of tools.
I wonder if the thinking budget can be optimized both for cost and speed. Any input on how to set the thinking budget would be great to get.
I have tried -1 as a dynamic budget value, which works fine but I don’t have a good intuition on what the benefits are for that.
We could run evals with different settings, but I was wondering if there are any heuristics or general advice that we could help getting started with this.