Most conversations about production AI agents focus on the agent itself — the prompts, the orchestration, the framework. But the moment you put an agent in front of real enterprise data, a different problem dominates: the data layer wasn't designed for this consumer. Data lakes were built for analysts and dashboards. Transactional systems were built for applications. Neither was built for a non-deterministic, token-hungry, latency-sensitive reasoning loop that may issue thousands of unpredictable queries per minute.
This talk takes the data architect's view of agentic systems. We'll walk through the architectural decisions that determine whether your agents are reliable and affordable in production — or quietly bankrupting your team.
You'll learn how to draw the line between deterministic and non-deterministic computation, and why getting that boundary right is the single biggest reliability lever you have. We'll cover when to route an agent to a transactional system versus a data lake, and what each choice costs in latency and consistency. You'll see how to add low-latency serving layers to a traditional data lake to make it agent-ready, how to design MCP servers that scale and stay secure under agent-driven traffic, and how a semantic layer can dramatically reduce hallucinations by translating raw schemas into agent-native concepts.
We'll also tackle the topic that quietly kills most agent projects: token economics. We'll share concrete patterns to keep agent workloads financially viable as they scale.
If you're moving agents from prototype to production, you'll leave with a decision framework — and a clear mental model for the data stack that has to exist beneath every reliable AI system.
Main Takeaways:
- Architectural strategies for serving data to AI agents — when to use transactional systems vs. data lakes, and how to add low-latency layers to make existing platforms agent-ready.
- Patterns for implementing scalable, secure, and token-efficient MCP servers under high-volume agent traffic.
- How a semantic data layer improves inference accuracy and reduces hallucinations.
Interview:
What is the focus of your work these days?
I am the Data Intelligence Director at Totvs, the largest tech company in Brazil. My mission is to provide data platforms and strategies for AI applications.
What is the motivation behind your talk?
As we deliver AI agents in production, several issues arise on how to provide data for AI agents in a secure and affordable way. I want to share what we learned in the last few years and provide a few techniques that can avoid common pitfalls.
What is your session about, and why is it important for senior software developers?
This talk focuses on the data layer — the part of agentic systems that quietly determines whether agents succeed or fail in production. Data lakes were built for analysts. Transactional systems were built for applications. Neither was designed for a non-deterministic reasoning loop issuing thousands of unpredictable queries per minute. Senior developers need a decision framework for this, not just another comparison of agent orchestration frameworks.
Why is it critical for software leaders to focus on this topic right now?
Because agents are moving from demo to production, and the real constraints are becoming visible. Precision, security, and cost are the three dimensions that determine whether your agent architecture is trustworthy and financially viable at scale. These aren't future concerns — teams running agents in production are hitting these walls right now, and the window to design the data layer correctly is before the technical debt locks in.
What are the common challenges developers and architects face in this area?
Four problems come up repeatedly: data staleness (agents confidently acting on outdated information), the read-versus-write split (data platforms are optimized for reads, but agents need to write back), scaling MCP tool catalogs without blowing up the context window, and keeping outputs precise and grounded when the underlying data has ambiguous or inconsistent definitions. Each has known solutions — integrating them into a coherent architecture is where most teams struggle.
What's one thing you hope attendees will implement immediately after your talk?
I hope attendees leave with a clear mental model for how to architect the data layer beneath an AI agent — which sources to use for which type of access, how to make existing data platforms agent-ready, and a concrete awareness of the pitfalls that kill projects in production: stale data, token bloat, prompt injection surface, and write-back consistency. The framework is straightforward once you see it; the hard part is knowing it exists before you've already made the wrong calls.
Who is your session for?
Data Engineers and AI Agents developers who need to deploy AI agents in production.
What makes QCon stand out as a conference for senior software professionals?
QCon consistently attracts practitioners who are working through hard problems in production, not just evaluating options in theory. The talks tend to be honest about failure, specific about tradeoffs, and free of the hype that dominates most AI conversations right now. For senior engineers who are past the "should we adopt AI" question and into the "how do we make this actually work" one, that calibration matters.
What was one interesting thing you learned from a previous QCon?
At QCon London 2026, I had the privilege of hosting the "Architecture in the Age of AI" track. The five talks were excellent, and I learned from every one of them. But the session that made me think differently was actually a keynote — a provocative talk on local-first computing. It challenged core assumptions about how we design software around data ownership and resilience. I left with a genuinely different mental model, which is exactly what you hope for from a keynote.
Speaker
Fabiane Nardon
Data Expert, Java Champion & Data Platform Director @totvs
Fabiane Bizinella Nardon is a tech executive with 20+ years architecting large-scale data systems. She currently leads Data Intelligence at TOTVS, Brazil's largest tech company, where her team designs data platforms and engineering strategies for the AI era — enabling LLM and agent-driven products with strong foundations in governance, security, and cost-efficient operation, including token usage optimization and trustworthy retrieval.
Previously, she was CTO at Tail (acquired by TOTVS in 2020), where she led data and ML systems processing 4B+ new records per day. Her PhD focused on Ontologies and Semantic Data Models, and she was an early practitioner of RDF and semantic technologies — foundations now critical for grounded AI systems.
Fabiane is a Java Champion and co-founder of SouJava. She hosted the Architecture in the Age of AI track at QCon London 2026 and has served on program committees for QCon London and QCon San Francisco.
