Umais Siddiqui's blog
- Oct 22, 2026
Let a script fetch the files, let the skill judge them
When an agent step is slow bulk fetching, move it into a one-time script and keep the skills for judgment, documentation, and assignment.
- Oct 22, 2026
Grade what the agent shipped, not the steps it took
AI agent evals should check the final output the user receives, because an agent can pass every step check and still ship something broken.