← Table of Contents · Page 684 of 868

PROJECT TWO - ADDING CONTEXT MANAGEMENT

Part 11 of 11

CHECK YOUR UNDERSTANDING • What is the difference between context bloat and context rot? ◦ Answer: Bloat means the context is larger than needed, which raises cost and delay, while rot means the model attends less reliably to facts inside a long context. • Why does the compaction function copy important messages word for word? ◦ Answer: Because a summary that paraphrases or drops a match decision could make the agent silently forget a result, so decisions must be preserved exactly. • How does the agent know after a restart which resumes are finished? ◦ Answer: It reads progress_notes.md, where the loop wrote one DONE line per finished resume, and skips those files. • Name two things a fair before and after test must hold constant or check. ◦ Answer: It must use the same tasks each time and change only one thing, and it must check quality as well as cost. • Why should you not treat the 83 percent figure as a promise for real systems? ◦ Answer: Because the token count is a rough estimate and the model is a scripted stand-in, so real results will vary and should be tested repeatedly.
Advertisement
Read Aloud:
0:00 / 0:00