Agents are data-hungry by design. Every piece of data they touch becomes a compliance liability if it is retained longer than necessary, stored unencrypted, or sent to a third-party model without appropriate controls.
LLMs perform better with more context. That creates a constant architectural
pressure to load everything available — full conversation history, broad
retrieval results, complete records rather than relevant fields. The
compliance answer runs directly against that pressure. You must design
retrieval to be task-scoped, not capability-maximizing.
This is not a guideline to follow loosely. Every major framework imposes the
same constraint from a different angle: