AWS shows market surveillance with LangGraph and Strands agents

Amazon's technical post demonstrates a production-ready architecture for multi-agent market surveillance systems built on two complementary frameworks: LangGraph and Strands Agent. The system orchestrates multiple specialized agents (security_monitor, broker_monitor, risk_monitor, intel_analyst) through a central workflow graph, each contributing its own insights to a final synthesizer node that produces comprehensive reports.

Strands Agent serves as the reasoning engine within individual nodes. The security_monitor agent in the example is configured with three tools: get_report_list (which loads available reports for a specific agent), get_report_schema (which retrieves column definitions and parameters for a report), and run_report (which executes a validated, parameterized SQL query). This separation of discovery from execution is intentional—it prevents the LLM from writing raw SQL and eliminates the risk of filter injection. The agent runs on Anthropic's Claude Sonnet 4.6 model via Amazon Bedrock, configured with a maximum of 16000 tokens, adaptive thinking with an 8000-token budget, and default prompt caching to manage cost and context window usage. Strands handles comprehensive session and state management, including multiple conversation managers, so the context window does not overflow during multi-step analysis.

LangGraph provides the macro-level orchestration layer. It models the workflow as a directed graph in which nodes represent functions containing agent logic and edges determine execution flow. The AgentState TypedDict maintains a shared state—including the query text, session ID, required agents list, and insights contributed by each specialist. After every node execution, LangGraph's checkpoint system snapshots the complete workflow state via AgentCoreMemorySaver, enabling recovery from failures and human-in-the-loop review. An orchestrator node routes incoming queries to the appropriate specialists based on the required_agents list; a route_analysts function then walks through each specialist by index, allowing conditional branching and dynamic routing. The graph includes built-in retry strategies with exponential backoff and comprehensive observability through OpenTelemetry for integration with most monitoring platforms.

The combination of both frameworks solves the core architectural challenge: enterprise workflows must be deterministic and auditable (following predefined compliance and routing logic), yet certain steps within those workflows require the flexible reasoning of an LLM. By placing Strands agents at specific nodes within the LangGraph graph, the system isolates LLM intelligence to exactly where it is needed. Each agent manages its own hyper-focused context and tool history, while LangGraph maintains the structured, overarching session state that all agents can reference and update independently. This compartmentalization reduces hallucination risk, improves security against prompt injection, and makes the system auditable—critical for financial services. The complete solution, including code samples and the full workflow diagram, is available on GitHub.

The post addresses a critical tension in enterprise AI: the need for deterministic, auditable workflows versus the flexibility of LLM reasoning. Traditional single-agent systems either lack the structure necessary for compliance and reliability, or they constrain reasoning too severely. LangGraph solves this by treating workflows as directed graphs with persistent state management and checkpoint-based recovery—features essential for production systems handling human-in-the-loop review and long-running investigations. Strands Agent, meanwhile, provides the reasoning engine within specific nodes, applying LLM intelligence only where ambiguity or dynamic analysis is required. This separation of concerns is particularly valuable for financial services, where workflows must follow regulatory logic (market surveillance, reporting, compliance gates) but still benefit from intelligent analysis of trading patterns and suspicious activity.

The architecture also addresses security and cost concerns. By having Strands agents explicitly retrieve data schemas before querying, the system avoids hallucinated SQL and injection attacks. The use of Anthropic's adaptive thinking (with a token budget) and prompt caching on Bedrock further controls cost and context management. The persistent state layer maintained by LangGraph ensures that if a node fails or a human analyst needs to review intermediate results, the system can resume from exactly where it left off—a critical capability for investigations that may span hours or days.

No comments yet. Be the first to share your thoughts!