AWS has introduced Deep Agents and Bedrock AgentCore, tools designed to help developers build context-rich research agents that can handle complex workflows with isolated execution environments. These tools enable teams to manage large language model (LLM) context windows more effectively by delegating deep work to isolated subagents that return only concise results. The solution is aimed at developers working on multi-step AI workflows that require isolated execution environments for their agents. The tools include a real browser in a MicroVM for web research and a full Python environment for data analysis. This approach helps teams avoid the limitations of manual prompt-chaining or sequential processing, offering a more efficient way to manage AI research tasks.
The new tools allow developers to create agents that can perform tasks like web research, data analysis, and report generation in parallel. For example, a coordinator agent can check past research insights, spawn browser subagents to research different competitors, and use an analyst subagent to generate comparison charts and markdown reports. Each subagent type accesses only its specific set of tools, ensuring isolation and efficiency. Developers can also trace the entire workflow using Amazon CloudWatch or LangSmith, providing visibility into the execution process. The tools are available as part of the Deep Agents CLI, allowing users to run AgentCore CodeInterpreter without building a full agent.
The source text provides an in-depth walkthrough of how to build the agent, including setting up the model, creating toolkits for each subagent type, and wiring them together with LangChain Deep Agents. It outlines the necessary prerequisites, such as an AWS account with Amazon Bedrock AgentCore access enabled, AWS credentials, and Python 3.11 or later. The implementation includes code examples for configuring the LLM, creating browser toolkits, and setting up the interpreter toolkit for data analysis. The solution also highlights optional features like cross-session memory and observability tools for tracking agent performance and insights.
Source: awsml
