Deep Agents + Bedrock AgentCore: context-rich research agents

LangChain Deep Agents and Amazon Bedrock AgentCore present a pattern for context-rich research agents that keeps deep browsing and data analysis out of the coordinator’s context window. The coordinator spawns ephemeral subagents: parallel browser subagents in AgentCore Browser MicroVMs and an analyst subagent that runs a Code Interpreter, then synthesizes concise findings and saves them to AgentCore Memory.

How does the architecture work?

The orchestrator stays lightweight while isolated subagents handle deep work: browser subagents run real Chromium inside AgentCore Browser MicroVMs, an analyst subagent runs AgentCore Code Interpreter in a separate MicroVM, and a coordinator uses AgentCore Memory for recall. The orchestrator is the LangChain Deep Agents component, the model example is ChatBedrockConverse configured with model "us.anthropic.claude-sonnet-4-6", and observability is provided by Amazon CloudWatch or LangSmith.

In practice the coordinator first checks AgentCore Memory for past insights, then spawns parallel researcher subagents, each with its own Browser MicroVM. Each browser MicroVM connects through Playwright and exposes tools such as navigate_browser, extract_text, click_element, type_text, scroll_page, extract_hyperlinks and wait_for_element. When research subagents return structured findings, an analyst subagent uses the Code Interpreter tools (execute_code, execute_command, write_files, read_files, list_files, upload_file and install_packages) to generate charts and markdown. The workflow emits traces and spans in OpenTelemetry format so you can inspect the coordinator, child spans for each research subagent, and the analyst span in CloudWatch GenAI Observability or via LangSmith tracing.

How do you build one step by step?

Set a Bedrock-backed model, provision browser toolkits for each researcher, add a Code Interpreter toolkit for analysis, and optionally wire AgentCore Memory for long-term recall. The example configures a ChatBedrockConverse model with model="us.anthropic.claude-sonnet-4-6" and creates one BrowserToolkit per competitor URL. The sample COMPETITORS list includes GitHub (https://github.com/pricing), GitLab (https://about.gitlab.com/pricing) and Bitbucket (https://www.atlassian.com/software/bitbucket/pricing).

Each browser toolkit creates an isolated MicroVM session. The example increases session_manager.session_wait_timeout to 60.0 seconds (default: 10 seconds) to allow concurrent browser operations to finish. The interpreter toolkit provisions a MicroVM with a full Python environment and packages such as pandas, matplotlib and numpy pre-installed. Memory is optional but supported: the sample uses a MemoryClient to create events; AgentCore Memory requires at least one extraction strategy so that saved events produce extracted knowledge for future recall. Finally, create_deep_agent wires the model, research and analyst subagents, and memory tools together; the notebook states expected runtime of 4–6 minutes with Claude Sonnet for the three-site scenario, with sequential processing taking up to 3x longer.

Why it matters

This pattern keeps the coordinator’s context focused on synthesis rather than raw source content, which avoids filling the model’s context window with long page dumps. Isolation also narrows failure modes: browser issues stay inside a browser subagent span and interpreter problems inside the analyst span, making debugging and tracing more targeted. Running three browser research tasks concurrently in separate MicroVMs reduces wall-clock time compared with sequential browsing, and the same AgentCore tools work regardless of which model you choose.

What to watch

Watch for Part 2 of the notebook, which deploys the same agent to Bedrock AgentCore Runtime so the agent runs as a managed, session-isolated service with OTEL instrumentation enabled automatically. Also check whether your account enables CloudWatch Transaction Search, required once to surface AgentCore traces and spans in CloudWatch GenAI Observability.