Market surveillance agent with LangGraph and Strands on AgentCore
AWS Gleb Geinke
AWS shows how to build a market surveillance AI system by pairing LangGraph with Strands agents on Bedrock AgentCore. It's a blueprint for mixing rigid, auditable workflows with flexible LLM reasoning where compliance really matters.
Based on reporting by AWS, Gleb Geinke — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Financial surveillance is one of those domains where you can't just point a chatbot at trading data and hope for the best. Regulators want predictable, auditable processes, but spotting suspicious trading patterns often requires the kind of flexible judgment that only a language model can bring. AWS's latest technical walkthrough tackles that tension head-on by combining two frameworks that solve different halves of the problem: LangGraph for the rigid, checkpointed workflow skeleton, and Strands for the reasoning that happens inside individual steps.
The architecture splits responsibilities cleanly. LangGraph acts as the orchestrator, modeling the whole investigation as a directed graph — an orchestrator node routes tasks to specialist agents like security_monitor, broker_monitor, risk_monitor, and intel_analyst, then hands off to a synthesizer that compiles the final report. Every node execution gets checkpointed automatically, which means a failed run doesn't wipe out hours of investigation work, and it also opens the door to human-in-the-loop review when an analyst needs to step in mid-process. That's not a small thing for compliance teams who need to show their work.
Inside each node, Strands does the actual thinking. AWS's example security_monitor agent runs on Claude Sonnet 4.6 through Amazon Bedrock, and it's built with a deliberately paranoid tool design: separate functions for discovering what reports exist (get_report_list, get_report_schema) versus actually pulling data (run_report). The run_report tool validates every filter against a known schema and builds parameterized SQL itself — the model never writes raw SQL. That's a direct defense against prompt injection and hallucinated queries, which matters a lot when the thing doing the reasoning is a probabilistic model with access to trading databases.
What's notable here isn't the novelty of either framework alone — LangGraph's checkpointing and Strands' model-agnostic tool loop have both been around — but the specific division of labor AWS is proposing: deterministic orchestration on the outside, contained bursts of LLM autonomy on the inside, each specialist agent getting its own isolated context so it doesn't drift or lose track of instructions. Deployment then runs through Bedrock AgentCore, which AWS pitches as handling the containerization, scaling, and session management so teams aren't stitching together their own infrastructure for long-running, stateful agent workflows.
The GitHub repo backing this post suggests AWS wants developers to actually build this, not just read about it — which fits a broader pattern of AWS using detailed reference architectures to push Bedrock adoption in regulated industries where "just trust the model" isn't going to fly with a compliance officer.
My take — AI-written commentary, not fact-checked reporting
This is a sensible pattern, and I'd bet we see more of it as enterprises realize pure agentic autonomy is a liability in anything regulated — banks, healthcare, insurance. What I find genuinely interesting is the SQL-injection defense baked into the tool design; that's the kind of unglamorous engineering that actually determines whether an LLM system survives contact with a real compliance audit, and it gets far less attention than whatever benchmark score the underlying model posted last week.
Read more about this at: AWS