At AGNTCon + MCPCon in San Jose, Jeu George, Co-Founder and CEO of Orkes, joins The New Stack’s Nick Lucchesi to discuss the challenges of building mission-critical agentic applications and the infrastructure needed to support them.

Building an AI agent that performs an impressive demonstration is one thing. Building one that can reliably process financial transactions, handle customer refunds, or make business decisions without compromising sensitive data is quite another.

That distinction is becoming increasingly important as enterprises move from experimenting with generative AI to deploying agentic applications in production. Developers need more than powerful language models. They need dependable access to enterprise data, orchestration, security, fault tolerance, and a way to explain what an AI agent actually did. Orkes believes its newly previewed ParticleDB can help solve a significant part of that problem.

In an interview last month with The New Stack Editor-in-Chief Nick Lucchesi at AgentCon and MCPCon 2026 in San Jose, California, Orkes co-founder and CEO Jeu George discussed how the company’s emerging database technology and established Conductor orchestration platform are designed to make enterprise AI development simpler and more dependable.

Why ParticleDB matters to developers

At its recent Orkes Shift developer conference in San Francisco, the company previewed ParticleDB, an agentic database designed to bring together data capabilities that conventionally require multiple systems. Enterprise developers frequently work with relational databases, key-value stores, vector databases, and analytical systems. Connecting these technologies often means building complex extract, transform, and load (ETL) pipelines, managing data movement, and maintaining infrastructure that becomes increasingly difficult to operate.

ParticleDB’s premise is that these capabilities should work together within a common engine, reducing the integration burden on development teams. George contends that the need becomes especially pressing when AI agents enter the picture.

“When do AI agents and LLM models become useful? When they have data.”
—Orkes CEO Jeu George

“When do AI agents and LLM models become useful? When they have data,” George told The New Stack. That data, however, often resides within tightly controlled enterprise environments. Companies cannot simply expose sensitive customer information, financial records, or proprietary business processes to autonomous agents.

George envisions a different approach: Rather than allowing agents to roam freely through enterprise systems, a trusted platform supplies the specific information and context they need. Combined with Orkes Conductor, ParticleDB is intended to keep applications, agents and their underlying data closely integrated within customers’ cloud boundaries.

Reliability is still the hardest problem

Beyond data integration, George emphasized a familiar challenge for distributed-systems developers: Things inevitably break. Network outages, software bugs, infrastructure failures, and security incidents are unavoidable. With AI agents potentially executing workflows for hours or even days, recovering from failure without duplicating transactions becomes essential.

“If you design systems from the ground up that things are always going to fail, you simulate failures all the time, and that’s how you get to a robust application.”

“If you design systems from the ground up that things are always going to fail, you simulate failures all the time, and that’s how you get to a robust application,” George said. Orkes Conductor addresses that problem through workflow orchestration, recovery mechanisms, security controls, and audit trails. Developers can combine deterministic business processes with AI-driven decisions while retaining visibility into individual actions.

For regulated industries such as banking and insurance, that visibility is particularly valuable. Organizations need to reconstruct an agent’s decisions, identify which models and systems were involved, and explain outcomes to customers or regulators.

Knowing when to use an agent

One particularly useful part of the conversation examines a straightforward e-commerce refund. George explains why AI agents should handle ambiguous tasks, such as interpreting customer complaints, while predictable operations – issuing refunds, updating financial records and generating receipts – remain conventional workflows. Unusual or high-risk cases can be escalated to humans.

“…successful agentic applications aren’t necessarily the most autonomous. They’re the ones that combine intelligence with dependable execution.”

The broader message for developers is that successful agentic applications aren’t necessarily the most autonomous. They’re the ones that combine intelligence with dependable execution.

Watch our full interview with Jeu George to learn how Orkes is approaching enterprise data integration, agent orchestration, security, and explainability – and why George believes the next generation of AI applications will require a fundamentally different production infrastructure.

The post How Orkes ParticleDB turns an impressive demo into a remarkable application appeared first on The New Stack.