Is cryptocurrency a good buy for Q2? Which supplier clears our compliance bar at an acceptable cost? Which biomarkers matter for early pancreatic cancer detection? Different fields, same shape of work each time: query several sources, cross-reference what comes back, and write it up. AgentCore Deep Research is an open-source application on Amazon Bedrock AgentCore that does that end to end, with inline citations and diagrams. It ships with over ten data connectors and a React frontend, and deploys as a single CDK stack.

The agent works in four steps, and their order is the interesting bit. It scaffolds first, reading the question, picking three themes and writing an empty report skeleton, before searching for anything at all. Then it runs four to six searches, judges source quality, and follows up wherever findings are thin or contradictory. It writes section by section, citing as it goes, then re-reads the whole report, checks each section actually says something, and fills the gaps it finds.

Committing to a structure before gathering evidence sounds backwards, but it stops the agent wandering. The skeleton gives it a definition of "done" that doesn't depend on what the search results happen to contain, and the verify pass exists for the same reason, since an agent writing sequentially will otherwise leave thin sections behind and never notice. If the report contains numbers, an optional step generates charts through AgentCore Code Interpreter, running matplotlib in a sandbox. It builds on the Fullstack AgentCore Solution Template.