SQLBits returned to the ICC Wales in Newport from 22 to 25 April for four days of solid content, good networking, and the usual conference energy. The cartoon theme was a nice touch. This post covers the three areas I found most valuable: AI agents and the Foundry stack, Model Context Protocol, and Polars as a genuine alternative to Spark on Fabric.

Orange SQLBits logo of the Company name in the font Anton.

AI Agents and Azure AI Foundry

AI and agentic AI were a dominant theme, with sessions covering LLMs, RAG, Copilot-style assistants, and agent-driven workflow automation. The standout for me was Microsoft Foundry Demystified: Building and Evaluating Agents with MCP and Guardrails. It walked through creating an MCP server as an Azure Function, connecting it to an agent, and using Foundry’s evaluation and guardrail features to keep automations safe and aligned with organisational requirements. Fast-paced, live demos, and a few specifics on guardrail configuration I had not seen documented clearly before.

A second session on building production-grade agents with Azure AI Foundry paired well with it. Together they gave a complete picture of the current Foundry agent story from first principles to deployment.

Azure AI foundry logo on a white background

Model Context Protocol

MCP Servers for Conversational BI made a solid case for MCP sitting between a language model and a BI layer to enable natural language interaction with data.

Synthetic Data Generation with AI was a 20-minute session that covered more ground than most hour-long talks. It compared approaches from simple VS Code and Fabric notebook setups through to dedicated MCP servers for more controlled, repeatable agentic creation workflows. The honest trade-off framing was refreshing: MCP servers give you more control but cost more to set up. Know which one your situation actually needs.

Seeing MCP get serious airtime at SQLBits confirms it has moved well past the experimental stage. Exposing enterprise tools and data via MCP to an agent layer is becoming a standard architectural pattern.

Polars on Fabric

Spark Unplugged: How In-Process Analytics Is Making Distributed Computing An Expensive Investment was one of the most practically useful sessions of the week. The benchmarks showed in-process Python engines (specifically Polars and DuckDB) running four to five times cheaper and faster than Spark for common workloads on Fabric. The case for Polars extended beyond raw speed, with benefits including faster test suites, lower platform costs, and an API that developers actually enjoy using. Real migration data from production workloads gave it genuine weight rather than a theoretical benchmark.

If your data fits in memory and you are defaulting to Spark out of habit, it is worth giving Polars a serious look. This session is worth watching when the recordings land.

Logo for Polars which is a striped polar bear icon

Final Thoughts

SQLBits 2026 felt grounded in a way the last few years have not always managed. Less hype, more working code. The ICC Wales works well as a venue and the community continues to be the best reason to go.

If you missed it, the recordings will be worth your time. See you there next year.

Tags: ,

« Previous Post
Next Post »