Define the task
I begin with a workflow from my work. Which sources must the tool read, what changes may it make and how will the result be checked? Only then can I assess whether it fits the task.
Consider operation
Quota and failures also matter in my agent platform, Harness. The router considers remaining allowance and previous failures. A model that is currently unavailable cannot help with the task at hand. Daily budgets limit paid calls.
Use the same sources
A comparison needs the same task and acceptance checks. I record what was actually checked. One useful run is a reason to use a tool again.
Separate tools from responsibility
Models and coding agents are the tools. I remain responsible for the architecture and rules. Whether a change is accepted depends on its checks.