LangGraphとStrandsを使ったAgentCore上のマーケット監視エージェントMarket surveillance agent with LangGraph and Strands on AgentCore

Amazon Web Services has published a technical walkthrough explaining how to build a production-ready agent">multi-agent AI system for market surveillance, pairing the LangGraph framework for workflow orchestration with the Strands library for agent reasoning, all running on Amazon Bedrock AgentCore. The guide matters because it tackles one of the persistent challenges in applied generative AI: moving agent">multi-agent designs from experimental notebooks into reliable, observable production systems, and it does so through a regulated, high-stakes domain that leaves little room for unpredictable behavior.

Market surveillance—the practice of monitoring trading activity to detect manipulation, fraud, and compliance breaches—is a fitting example. It typically involves ingesting large volumes of market data, applying layered analytical reasoning, and producing decisions that must be auditable. A single monolithic prompt is poorly suited to that kind of work, which is why the post structures the solution as multiple specialized agents coordinated by an explicit workflow rather than a single model call.

In the described architecture, the two frameworks play complementary roles. LangGraph manages orchestration through what AWS calls state-driven workflows, modeling the process as a graph of steps in which transitions depend on shared state rather than a fixed linear script. Strands, meanwhile, handles the reasoning inside individual agents, giving each agent the logic it needs to interpret inputs and decide on actions. Separating orchestration from reasoning in this way is a common pattern for keeping complex agent systems maintainable, since the control flow can be reasoned about independently of any one agent's internal behavior.

A central technical detail is checkpointing. The post highlights checkpoint-based mechanisms that persist the workflow's state at defined points, which allows a run to be paused, resumed, or recovered after a failure without starting over. In practice this supports resilience for long-running processes and can enable human-in-the-loop review, where an analyst inspects or approves intermediate results before the workflow continues. For a compliance-oriented use case, the ability to capture and replay state also contributes to traceability.

The deployment target, Amazon Bedrock AgentCore, is AWS's platform for running agents in production. AgentCore is positioned as a set of services—covering areas such as runtime execution, memory, identity, tool access, and observability—that aim to handle the operational concerns of hosting agents at scale. Building on AgentCore appears intended to let teams focus on agent and workflow logic while relying on managed infrastructure for scaling and monitoring, though the specific services and configurations used are detailed in the original post.

Some background helps situate the pieces. LangGraph is an open-source library from the team behind LangChain, designed for building stateful, multi-actor applications with language models by expressing them as graphs. Strands Agents is an open-source, model-driven agent framework that AWS has promoted for building agents with relatively little boilerplate. The combination is notable because it mixes a third-party orchestration framework with AWS's own tooling on AWS infrastructure, reflecting a broader industry trend toward interoperable, framework-agnostic agent stacks rather than a single vendor's end-to-end product.

That trend has accelerated as major providers race to offer agent-building tools. Amazon Bedrock itself provides access to a range of foundation models and previously introduced Bedrock Agents, while competing ecosystems from other cloud and model vendors have released their own orchestration and agent frameworks. Emerging conventions for connecting agents to tools and data, and for agent-to-agent communication, are also shaping how these systems are assembled. AWS's decision to demonstrate LangGraph alongside Strands is consistent with that more modular direction.

For teams evaluating the approach, the walkthrough is best read as a reference architecture rather than a turnkey product. Prerequisite concepts include familiarity with agent">multi-agent orchestration, graph-based state machines, and the operational requirements of running LLM-driven systems, such as latency, cost, and error handling. Readers considering the pattern for regulated workloads would likely need to add their own controls around data governance, model evaluation, and auditability beyond what a sample implementation covers. As a demonstration, the post illustrates how existing open-source frameworks and AWS's managed agent platform can be combined to address orchestration, reasoning, and resilience together, using market surveillance as a concrete and demanding scenario.