Paper 2 · Question 6

A: Scratchpad files offload intermediate results from the context window. Each step writes its output to a file, keeping the context budget free for the current step's reasoning. The final step reads back only the aggregated data it needs, avoiding the accumulation of raw intermediate results in context.

B: Cross-database joins between Snowflake and PostgreSQL are not possible in a single SQL statement — they are separate database systems accessed through different MCP tools. The multi-step approach is architecturally necessary.

C: Subagent delegation is useful for independent parallel tasks, but this analysis has sequential dependencies (each step depends on the previous step's output). Subagents would need the same intermediate data passed to them, merely shifting the context problem rather than solving it.

D: Summarising intermediate numerical results risks the progressive summarisation trap — exact billing figures, usage counts, and exchange rates may be compressed into qualitative descriptions, causing calculation errors in subsequent steps.