AWS Launches Deep Agents Framework with Bedrock AgentCore for Isolated Multi-Agent Workflows
AWS has launched LangChain Deep Agents, an orchestration framework that spawns isolated subagents using Amazon Bedrock AgentCore infrastructure. AgentCore provides dedicated browser MicroVMs running Chromium via Playwright and full Python environments with pre-installed data science libraries, enabling parallel agent execution without context window competition.
AWS Launches Deep Agents Framework with Bedrock AgentCore for Isolated Multi-Agent Workflows
AWS has released LangChain Deep Agents, an orchestration framework that addresses context window limitations in multi-step AI workflows by delegating tasks to isolated subagents. The framework pairs with Amazon Bedrock AgentCore, which provides dedicated execution environments including browser MicroVMs and Python interpreters.
Core Infrastructure
Bedrock AgentCore provisions ephemeral MicroVMs for each subagent type. Browser subagents receive isolated Chromium instances connected via WebSocket using Playwright, with tools including navigate_browser, extract_text, click_element, type_text, scroll_page, extract_hyperlinks, and wait_for_element. Default session timeout is 10 seconds, configurable up to 60 seconds for concurrent operations.
Code Interpreter subagents run in separate MicroVMs with Python environments pre-configured with pandas, matplotlib, and numpy. Available tools include execute_code, execute_command, write_files, read_files, list_files, upload_file, and install_packages for runtime library additions.
Architecture Pattern
The framework implements a coordinator-subagent pattern. A coordinator agent receives requests, spawns specialized ephemeral subagents for parallel execution, and aggregates results. Each subagent type accesses only its specific toolkit, preventing tool sprawl and context contamination.
In AWS's reference implementation, a coordinator spawns three browser subagents in parallel for competitive research, each navigating different websites in isolated MicroVMs. An analyst subagent then receives combined findings and generates comparison charts using the Code Interpreter. The coordinator can optionally save insights to AgentCore Memory for cross-session retrieval.
Memory and Observability
AgentCore Memory provides long-term knowledge storage with extraction strategies that automatically structure insights from conversation events. According to AWS, Memory resources require at least one configured extraction strategy (such as semanticMemoryStrategy) for recall functionality to work—without strategies, only raw events are stored.
Tracing integrates with Amazon CloudWatch through AgentCore Observability or LangSmith. Each subagent execution generates distinct session IDs, confirming isolation.
Deployment Options
Developers can run Deep Agents locally with deepagents --sandbox agentcore to test AgentCore CodeInterpreter without building a complete agent. AWS also provides an AgentCore CLI for deploying agents to AgentCore Runtime as managed, session-isolated services.
The framework supports Claude Sonnet through Amazon Bedrock, using cross-region inference profiles with the us. prefix for higher availability. Sessions spin up in seconds according to AWS documentation.
Requirements
Prerequisites include an AWS account with Bedrock AgentCore access, configured AWS credentials via environment variables or CLI profile, and Python 3.11 or later. IAM permissions are detailed in AWS's AgentCore getting-started guide.
What This Means
This release directly addresses the context window limitation in multi-step agent workflows where raw content, code execution, and reasoning compete for token space. By isolating execution in dedicated MicroVMs, Deep Agents enables genuinely parallel agent operations without manual prompt-chaining. The browser MicroVM approach is particularly significant—most agent frameworks rely on HTML extraction APIs rather than full browser automation, limiting their ability to interact with dynamic web applications. AgentCore Memory's automatic extraction strategies could reduce redundant research across sessions, though effectiveness depends on extraction strategy configuration quality.
Related Articles
Guardoc Health Cuts Documentation Errors 46% Using Amazon Nova Models on Bedrock
Guardoc Health built a multi-stage document processing pipeline on Amazon Nova Pro, Nova Lite, and Titan Text Embeddings to extract and classify medical conditions from clinical PDFs at scale. The company claims a 46 percent reduction in documentation errors, 70 percent fewer audit fines, and over $400K in annual ROI for a single facility.
Cursor's New Agent Swarm Architecture Splits Planning and Coding to Cut Costs by Up to 15x
Cursor's upgraded agent swarm splits AI coding work between expensive frontier planner models and cheap worker models, reaching 100 percent accuracy on a from-scratch SQLite-to-Rust benchmark. Cursor claims the split cuts codebase size by up to 85 percent and worker costs by a factor of 15 compared to running a single frontier model alone.
GitHub Copilot App Adds Stacked Sessions and Pull Requests for Sequential Coding Tasks
GitHub has published a walkthrough of stacked sessions and pull requests in the Copilot app, a workflow feature that lets developers chain multiple AI-assisted coding sessions and PRs together. The post uses a real-world example: modernizing an old codebase.
GitHub Copilot App Adds Stacked Sessions to Chain Multiple AI Coding Tasks Into Sequential Pull Requests
GitHub has added stacked sessions and pull requests to the GitHub Copilot app, allowing developers to break large coding tasks into a sequence of dependent Copilot sessions that each produce their own PR. GitHub demonstrated the feature by using it to modernize an old codebase.
Comments
Loading...