Autonomous systems force software architects to answer a question that traditional applications rarely had to ask: where should analytical execution happen?

As organizations deploy more agents into production, many drift toward one of two extremes. Either agents receive increasingly broad access to operational systems, increasing governance, operational, and cost risks, or each new agent acquires its own retrieval pipeline, vector store, cache, and data infrastructure, creating duplicated systems, stale copies, fragmented ownership, and growing operational complexity.

This session examines the architectural decision that sits between those extremes. Using the same production agent workflow throughout, we will compare the principal execution models available today, including embedded in-process execution, direct operational database access, dedicated data services and APIs, central warehouses and lakehouses, and federated query layers.

Rather than asking which architecture is best, we will examine the engineering forces that lead organizations toward each approach, the workloads where each performs well, and the operational constraints that eventually limit each one. Identity, provenance, authority, replayability, and operational ownership are introduced only where they help explain production failures, including agents with excessive authority, per-agent data stacks serving stale information, and federated queries that could no longer be audited.

Within the first part of the session, attendees will see the same data-intensive agent task implemented across three execution models, comparing execution paths, latency, freshness, and failure behaviour before developing a framework for selecting the most appropriate architecture.

Attendees will leave with a practical framework for deciding where analytical execution belongs within autonomous systems, understanding the trade-offs between different execution models, and recognizing when an architecture that works today may become tomorrow's production bottleneck.

What You Will Learn:

• How to evaluate different execution models for data access in autonomous systems
• The trade-offs between embedded execution, direct database access, services, warehouses, and federated query layers
• How architectural choices affect latency, freshness, operational complexity, authority, provenance, and failure behaviour

Who Should Attend:

• Software architects
• AI engineers
• Data platform engineers
• Platform engineers
• Staff and principal engineers
• Technical leads designing autonomous systems
• Engineers responsible for data platforms and AI infrastructure