Comment by slowin

I don't think agents are append only. At the end of the day, you're just presenting context to the LLM. That context can be pruned and compacted (and is). There's no guarantee that an iteration of an agent loop contains all prior context unmodified.