Reference: Deep Browser Research Agent

Architectural notes

Reference architecture for a deep-research agent that browses many
sources and synthesizes a structured report.

Why these picks:
- LangGraph for explicit checkpointing across long-running research
  flows (rate limits, retries, partial state recovery).
- Browserbase for fleet management of headless browsers at concurrency.
- Opus for the final synthesis step where reasoning quality dominates.
- Modal for compute scaling without committing to provisioned servers.

Common gotchas:
- Source quality dominates output quality; aggressive de-duplication
  helps more than better synthesis prompts.
- Long flows need explicit budget caps or costs run away fast.