agents
Why pay an agent to run a command you already know?
Asking an agent to list files takes longer than typing ls and costs tokens, only for the agent to run ls anyway. On paper, delegating a command you already know makes no sense.
An eval score is only half the result
Your agent's score improved. Its inference bill increased 15 times. Before you celebrate, decide what the improvement is actually worth.
Your agent eval is more than its result
Your agent completed the task and every grader passed. What happened before the score reached 100%?
Is your eval lying to you?
Your eval passes on every commit. But what did its graders actually prove?
A knowledge cutoff is a ceiling, not a snapshot
A model's knowledge cutoff looks precise. But how much does that date actually tell you about what it knows?