Nomura frames graph engineering as a safer path for enterprise AI agents

AI-powered systems are moving through an architectural progression from the Model Context Protocol to loop engineering and graph engineering. MCP gives models a standardized way to access tools, allowing financial services teams to connect agents to market data, credit ratings, regulatory filings and risk databases through a common, auditable interface.

Loop-based agents can pursue goals across multiple steps by thinking, acting, observing and repeating. That autonomy creates a control problem in regulated settings because the agent can determine when to stop, which may be unsuitable for trading desks and other high-risk workflows. Graph engineering adds structure by placing agents inside workflows with gates, checkpoints, approval steps and failure isolation.

Nomura presents graph engineering as especially relevant for capital markets, front-office trading, risk and compliance teams. The approach separates stable organizational graphs, which define roles, permissions and controls, from work graphs, which manage task execution and adapt workflows as evidence is gathered. LangGraph is described as the most widely adopted framework, with more than 65 million monthly downloads, offering checkpointing and human-in-the-loop support for governed agent systems.