On a crisp November morning, long before the city fully came alive, we gathered inside Databricks’ Amsterdam HQ for the first Builders CTO Breakfast, a quiet room filled with engineering leads who spend their days (and too many nights) wrestling with the edge of what AI can do in production. The setup was intentionally simple: strong coffee, warm pastries, and a handful of prompts designed to provoke real, technical conversations. No demos, no pitches, no slides. Just CTOs and founding engineers comparing the realities behind the choices they make every day.

Setting the room

As people arrived, the contrast became instantly visible. Everyone in the room personally uses AI tools - Cursor, ChatGPT, Copilot, Replit Agents - not as toys, but as extensions of how they write, reason, and ship. Yet when the conversation shifted toward organisational adoption, the tone changed. Enterprises move slowly. Procurement hesitates. Security teams stall. Even €20-per-seat licenses get second-guessed. And quietly, something more structural is happening: AI spend is starting to take the place of junior and mid-level hiring. It wasn’t said dramatically — it was said as a calm observation. Seniors are getting amplified; juniors are getting squeezed. The room understood this instantly.

The instability beneath every stack

The first prompt of the morning unlocked a conversation that could have lasted hours.

“If each model update breaks your product, is the problem the model - or your architecture?”

Everyone laughed, because everyone had lived it. Model drift is now a constant, system prompts behaving differently after updates, frameworks changing faster than teams can document them, vendor-specific abstractions adding more instability than clarity. What the room circled around was a shared truth: the only stable surface you own is your state, your memory, your tools, your permissions, and your evaluators. Everything below that line - models, frameworks, prompting techniques - is moving ground. Build too tightly on the unstable layer, and you live in crisis mode. Build on the stable one, and model churn becomes a manageable variable, not a fire drill.

Mega-agent or mesh? The real architectural divide

One of the liveliest parts of the morning was a debate that every engineering leader in the room has wrestled with: do we build a single, all-encompassing mega-agent, or do we build a constellation of smaller, specialised ones? The idea of “one assistant that does everything” has obvious appeal - one identity, one interface, one mental model. But as people dug deeper, the weaknesses surfaced quickly: brittleness at scale, opaque failure modes, and the challenge of debugging a system that makes decisions you can’t fully trace.

The alternative, a mesh of specialised agents, felt more honest to the complexity real companies operate in. Teams today already work this way: onboarding processes, support flows, sales ops, QA, ingestion, document processing. Each is a different domain with different rules. Specialised agents fit that reality. The challenge shifts from capability to coordination. Routing, arbitration, and orchestration become the real engineering problem. The room found a middle ground: a top-level router to own the flow, and a network of agents to own the expertise. In other words: the architecture of a real team, not a magic generalist.

You don’t really know until you ship

The most candid conversations came around observability. Tjadi (CTO at Billy Grace) framed it perfectly:

“You don’t actually know how an agent behaves until it’s in the hands of a real user.”

And it hit everyone the same way, because they’d all been there. Offline evals give a sense of comfort that evaporates the moment an agent touches messy real-world data. Reasoning shifts. Edge cases appear. Weirdness emerges. The group agreed that observability is no longer a layer you add at the end; it’s the foundation. Thinking traces, action logs, audit trails, shadow deployments, deterministic envelopes around stochastic behavior. These are the patterns teams are adopting not because they want to, but because not having them feels like flying blind.

Patterns of pain every CTO recognised

Across all discussions, similar pain points kept resurfacing. Memory is still a moving target, too much context becomes noise, too little becomes blindness, and everyone is still figuring out when and how to clean up agent memory without breaking behaviour. Permissions came up repeatedly: human role-based access simply doesn’t translate cleanly to agent actions. Teams are realising they need to map actions to privileges, not identities to roles.

Safety, too, was a dominant theme. Outbound actions - sending emails, creating tickets, updating systems - carry real risk. Silent failures are even scarier. And underneath it all, every CTO is trying to manage inference costs while avoiding getting locked into a single vendor, even though every company in the room inevitably relies on at least one closed model today.

The shadow of enterprise requirements

Even in a heavily technical room, security and data residency came up with surprising consistency. European data at rest is now a hard requirement, not a preference. Customers want reproducibility, audit logs, and clear decision histories. And perhaps the most interesting shift: for enterprise buyers, data stewardship is beginning to outweigh raw model performance.

A morning of honest engineering

The beauty of the breakfast format was how naturally the room shaped the conversation. People appreciated that Databricks stayed technical - no product demos, no marketing slides - just genuine discussion about architectural trade-offs. The roundtable rotation kept conversations sharp, and by the end, several attendees asked for deeper sessions dedicated entirely to memory engineering, real-world observability stacks, orchestration patterns, and guardrail design.

The atmosphere felt a lot like the early Investor House editions, intimate, founder-driven, and refreshingly unpolished. A group of people who build for a living, finally in a room where they could drop the theoretical debate and talk about what actually breaks in production.

Closing

This Builders CTO Breakfast made one thing clear: agentic systems aren’t just a new technology layer, they’re forcing a new baseline for technical architecture. Memory is becoming the competitive differentiator. Governance can’t be bolted on later. Hybrid agent systems are becoming the norm. And observability is no longer optional if you want to sleep at night.

A big thank-you to the Databricks team for hosting us, to the builders who came ready to share openly, and to Tjadi and the Billy Grace team for grounding the discussion in real production truth.

Here’s to the next breakfast.