Designing the Handoff Between a Research Agent and a Writing Agent

You're building a system where a "research" agent gathers facts about a topic and hands off to a "writing" agent that drafts a report section from those facts. In an early version, the writing agent receives the research agent's entire tool-call transcript (every search query, every page it opened, every rejected source) and its output quality is inconsistent — sometimes it writes a good section, sometimes it misses the most important finding entirely.

1. Diagnose why passing the full transcript produces inconsistent output.
2. Redesign the handoff: what exactly should cross from the research agent to the writing agent, and in what form?
3. How would you tell, empirically, whether your redesigned handoff actually fixed the problem?