What practitioners learned in the room.

The FactSet session described a content-first MCP architecture that keeps routing, permissions, validation, and versioning behind a small set of self-describing tools. That makes institutional data easier for agents to consume without pushing FactSet's internal complexity into every client workflow.

The Deutsche Bank speaker argued that reliable agentic research starts with infrastructure, not prompts. Data must be discoverable, code reusable, and knowledge structured, controlled, monitored, and evaluated; otherwise an agent only reaches the wrong answer faster.

Databricks demonstrated a research-to-execution workflow on one governed lakehouse: structured and unstructured data, a multi-turn research agent, portfolio analysis, and factor implementation ran on the same platform without a separate middleware layer.

The dbt session argued that AI reliability depends on tested data contracts, shared metric definitions, and code lineage. Through dbt's MCP server, an agent could inspect models and provenance before producing a credit-risk dashboard and a Basel II audit-readiness report.

The QuestDB presenter argued for standards, open protocols and concise machine-readable documentation when building for coding agents. QuestDB paired that accessibility with billion-row financial analytics, allowing an agent to assemble a live ingestion and dashboard workflow from an empty project.

The market-data panel described stronger current results from automation, search, data work and code than from investment prediction. Scaling those gains requires governed data, traceability, realistic production economics, and enough trust to delegate a workflow without hiding its evidence.

The New York panel argued that finance is moving from interface-led analytics to API-first, agentic workflows. The durable advantage is shifting away from model choice and toward structured data, interoperability, domain expertise, and evaluation systems that make automated decisions trustworthy.