Build context-rich research agents with Deep Agents and Bedrock AgentCore
AWS shows how to build research agents that stay accurate without overloading a model's context window, using LangChain Deep Agents together with Amazon Bedrock AgentCore. The core trick is delegation: specialized subagents do isolated work and return only summaries, so the main model never drowns in raw content.
Key Takeaways
- The central idea is to keep the orchestrator's context clean by sending raw research to specialized subagents that return only summaries.
- Subagents run in isolation in separate MicroVMs, which speeds work and prevents tools from being misused.
- Parallel execution is much faster: a competitive-research workflow runs in 4 to 6 minutes, up to 3 times faster than running it sequentially.
- The design is model-agnostic, so teams swap Claude, Gemini, or another model with a single-line change.
- AgentCore handles deployment, memory across sessions, and observability through CloudWatch and LangSmith.
Stats & Key Facts
- #4 to 6 minutes runtime for the competitive-research workflow with Claude Sonnet
- #Up to 3x slower if the same research runs sequentially
- #Browser sessions auto-expire after 1 hour; Code Interpreter sessions after 15 minutes
- #AgentCore Runtime supports sessions up to 8 hours on ARM64 containers
The problem: depth versus the context window
Good research needs a lot of source material, but stuffing all of it into one model degrades quality. The architecture resolves that tension.
- ›Instead of flooding the orchestrator with raw pages, subagents read sources and hand back concise summaries.
- ›This keeps the main model focused and within its context limits.
- ›It mirrors how a human research lead delegates reading and asks for briefings.
How the architecture works
The system is a coordinator plus specialized workers, each with a narrow toolset.
- ›Three parallel browser subagents research competitor sites, each in its own MicroVM.
- ›A Code Interpreter subagent with pandas, matplotlib, and numpy builds the comparisons.
- ›A coordinator agent orchestrates the work through LangChain Deep Agents, and each subagent can only touch its own tools.
Deployment, memory, and model choice
Beyond speed, the platform covers the operational details of running agents in production.
- ›AgentCore Memory retains knowledge across sessions through extraction strategies.
- ›A four-step CLI flow moves an agent from notebook to managed endpoint with per-session isolation.
- ›Swapping the underlying model takes a single-line change, keeping the design vendor-flexible.
Frequently Asked Questions
What problem does this solve?
It balances research depth against a model's context limits. Rather than overloading one model with raw content, specialized subagents return summaries so the orchestrator stays accurate.
Why use separate subagents instead of one model?
Isolation makes the work parallel and faster, keeps each agent's tools scoped so they cannot be misused, and protects the main context from clutter.
How much faster is it?
The competitive-research workflow runs in 4 to 6 minutes, up to 3 times faster than doing the same work sequentially.
Can it use models other than Claude?
Yes. The design is model-agnostic, so teams can switch between Claude, Gemini, or other models with a single-line change.
The pattern is a practical template for production research agents: delegate to isolated, parallel subagents, summarize aggressively, and keep the model layer swappable.
Continue Learning
Comments
Sign in to join the conversation