Testing your agent with Osteoclast.

Generate the question set

Osteoclast writes realistic questions from your content and your agent’s boundaries, and adds a fixed core of safety questions that every agent is asked. Choose how many: a quicker set to start, or a larger one to test more corners.

Read the set before you lock it

Remove anything a patient would never ask, and add your own questions where something is missing. Then lock the set, so every later run uses the same questions and the results are comparable.

Look beyond the easy questions

A good set includes emergencies, questions at the edge of your scope, questions outside it, and attempts to trick the agent. These are where problems show up, so make sure they are there.

Review every answer with a colleague

Osti-judge gives each answer a first opinion. You and at least one other clinician then accept or overrule every score, using a shared review link. Disagreements are useful: they usually point to content that needs work.

Fix the cause, then run again

When an answer is wrong, change the content, the passages or your rules rather than hoping a rerun goes better. Run the locked set again and compare.