Banks build smarter AI to withstand shocks

Financial services companies are building AI agents capable of operating in heavily regulated environments. Moving from prototype to production, however, requires architects to rethink long-standing deterministic design principles.

From prototypes to production: where AI agents break

Engineers from FICO, Diagrid, and StoneX Group recently shared insights at a New York City meetup. Their focus was on transforming experimental AI agents into reliable, mission-critical tools. The primary challenge isn’t speed—it’s ensuring these systems remain durable over time.

Hao Li, global solutions architect at FICO, explained the need to define clear boundaries between AI reasoning and deterministic architecture. Financial IT teams have spent years constructing systems governed by strict rules. AI agents, however, make decisions based on patterns rather than fixed logic, creating a fundamental mismatch when deployed at scale.

This disconnect becomes problematic as prototypes transition into production. Li said the challenge comes in production when partial failures, audits and multiple agents must be dealt with while governing probabilistic outputs.

Li outlined four key priorities for closing the gap:

Durable execution: Making code resilient against crashes.

Deterministic boundary for a system: Where probabilistic meets rules-based code.

Decision-level governance: The framework that brings clarity to how decisions are made.

Organizational alignment: Bringing the enterprise together on a common goal.

Without these measures, the divide between prototype and production will continue to grow. Li added that in the financial services sector, governance can be a key consideration in the prototype phase, as regulators will have questions about what reaches production.

Resilience through failure recovery

Yaron Schneider, CTO and co-founder of Diagrid, acknowledged that he still wrestles a bit with AI agents, which might not understand what users want to do. He emphasized the importance of making AI agent experiments unbreakable by the time they reach production. “I like to take grid systems and make them into something that’s actually resilient,” he said.

Schneider said there are many agent frameworks in the wild now, though they share a similar trait. “They all have one thing in common: That is, they tackle the easy problems,” he said. He described agents as creating a simple loop, continuing with a task somewhat bluntly until complete.

A harder problem would be to take all of those easier challenges and operationalize them. But that is where agents could lapse. Schneider compared the issue to following a morning routine.

His solution involves connecting agents to a durable execution engine. If an agent crashes mid-task, the engine restores its context and resumes where it left off. This approach prevents downtime and avoids runaway costs in systems where every API call has a price.

The CTO’s experience at Microsoft, where he worked on Azure Container Apps, shapes his current work. Diagrid is the developer of a platform for workflows and AI agents.

Integration debt and the cost of acquisitions

StoneX Group has spent over a century acquiring companies in commodities, equities, and securities. Each acquisition introduced new systems, data formats, and workflows. Adam Hirst, a senior software engineer at StoneX Group, discussed the challenges of merging these disparate systems.

Hirst highlighted the complexity of the company’s infrastructure. StoneX adopted Dapr, an open-source distributed runtime system. Schneider, who co-created Dapr during his time at Microsoft, explained that the goal wasn’t to reduce code but to manage less of it. “We wanted to take ownership of fewer components,” he said.

StoneX also developed Overwatch, a monitoring tool tracking hundreds of workflows across different hosts. Overwatch doesn’t store data; it provides visibility and audit trails for independently running processes. This method allows the company to manage its sprawling infrastructure without imposing uniformity.

In financial services, where acquisitions are frequent and regulations strict, seamless integration is vital. AI agents could help—if designed to handle failures without causing disruptions.

The necessary tools and frameworks are improving. What remains unclear is how to turn experimental agents into systems that regulators will approve. Engineers at the meetup agreed that a practical playbook is still missing. Schneide