Build context-rich research agents with Deep Agents and Bedrock AgentCore
TL;DR
AWS walks through a competitive research agent that combines LangChain Deep Agents with Amazon Bedrock AgentCore. A coordinator delegates research, analysis, and synthesis to specialized subagents. Three browser subagents research GitHub, GitLab, and Bitbucket in parallel, each inside its own AgentCore Browser MicroVM. An analyst subagent uses AgentCore Code Interpreter to create a chart and markdown report.
Nauti's Take
This is one of the more useful AgentCore posts because it shows what a production-style agent can look like: browser isolation, code interpreter, memory, observability, and runtime deployment in one flow. It is still a vendor how-to.
Anyone rebuilding it should first check whether MicroVM costs, IAM complexity, and debugging overhead fit the use case. For real research agents, the practical lesson is clear: the coordinator should reason, the subagents should do the work.
Briefingshow
For developers, the important part is less the pricing-comparison demo and more the architecture behind it. Long agent workflows often become messy when research, raw content, code execution, and synthesis compete for the same context window. The AWS pattern separates that work: subagents do deep tasks in isolation and return compact results to the coordinator.
Sources
