Topic 4: Context efficiency and evaluation

Every tool definition you expose is sent to the model on every round, and every tool result stays in the conversation until the end. Context efficiency is the discipline of sending the model what it needs and nothing else. You cannot manage what you do not measure, so this part ends with an eval harness and a regression test.

Tool definition bloat and dynamic exposure

Tool definition bloat is what happens when a host connects to more servers than a question needs: every definition costs tokens per round, and every extra tool is another wrong answer the model can pick. Dynamic exposure means choosing which tools to send per request instead of sending all of them.

We compare four setups on the same 16 labelled questions (the eval cases defined in the next section) with the KeywordAgent stand-in: