Innovation Matters
Essays on AI, developer tools, and thoughtful software.
How do you evaluate an agent skill without a right answer?
A brainstorming skill might take five turns or fifteen and produce three ideas or ten. With no correct output to check, how would you know whether it helps?
Would you check an agent skill's nutritional label before installing it?
An agent skill for SharePoint developers now ships with a label that shows what it does to seven different agent setups. Would you want to see one on every skill you install?
What if your eval rubric is the problem?
After we updated an agent skill, our evals showed the same success rate at nearly twice the cost. Reverting the change looked like the obvious decision. So why didn't we?
I let LLMs play a 1988 detective gamebook 646 times
In a hot-air balloon with a storm coming, almost every model made the sensible choice and landed. About ten decisions later, that choice cost them the case.
You can't generate evals on autopilot
Writing eval criteria is hard, so you ask an LLM to write them. Minutes later, you have hundreds of reasonable-looking rubrics. But whose idea of a correct answer are they measuring?