[QUESTION]: Clarification on why gemini-fullstack mix direct and indirect state injection

[ykvidie]

Hi,

While reviewing the gemini-fullstack sample, I noticed that some agents (e.g. report_composer) use direct state injection in their instructions (e.g. {research_plan}, {section_research_findings}), while others (e.g. section_planner, section_researcher) only mention the state key in natural language instead of injecting it.

Could you clarify the design reasoning behind this mix?

• When should we prefer direct injection vs. relying on indirect context/state access?
• Is this mainly for prompt size control, or are there other architectural considerations?

This would help developers understand best practices when building their own workflows.

Thanks!