Agents & Data Engineering: A Community Night at Google
From buzzwords to the infrastructure that makes agents work
Last week, Data & AI Stockholm co-hosted an in-person evening at Google together with AI Engineering in Stockholm, bringing two communities into the same room around a shared question many teams are actively wrestling with:
What do we really mean when we talk about agents, and what needs to be in place for them to work in practice?
These kinds of sessions are exactly where our community shines. Different perspectives, different roles and experiences, but a shared interest in moving beyond buzzwords and into real engineering and decision-making.
We are grateful to Google for hosting the evening and opening up their space, and to everyone involved on both sides who helped make the night happen. A lot of coordination happens behind the scenes for events like this, but the quality of the discussions made it more than worth it.
From pipelines to context
The evening opened with Jim Dowling, Co-founder and CEO of Hopsworks, who grounded the conversation in the reality most data teams live in today.
“We are no longer just designing data pipelines. We are designing context.”
Modern systems are layered and imperfect by default. Batch and streaming coexist. Feature stores live alongside ad-hoc context. Inference pipelines rely on far more than model weights alone. Seeing this complexity visualised clearly helped anchor the conversation in what actually needs to work before agents can do anything meaningful.
One theme that resonated strongly was that seemingly “clean” agent behaviour often depends on a large amount of engineering underneath. Reusable context, evaluation loops, and disciplined iteration are becoming as important as model choice itself. This way of thinking is increasingly shaping how entire data teams work, not just ML engineers.
Agency as an economic shift
The evening concluded with Binny Francis, Head of Data for Telecom and Retail at Google Cloud, who offered both historical context and a practical lens on the evolution of agency.
Rather than treating agents as something entirely new, he placed them in a longer lineage of systems that act on our behalf. What has changed is the economics. As computing becomes cheaper, we increasingly trade a bit of precision for speed and scale, and that trade-off reshapes how systems are designed and trusted.
We also got a look into how this thinking shows up inside Google’s own data stack, from semantic metadata services to multimodal analysis directly in BigQuery, and a broader move away from static data layers toward more active, meaning-aware systems.
Why community events like this matter
What stood out most was not a single tool or framework, but the quality of the questions in the room:
How do you evaluate agent behaviour over time?
Where does responsibility sit when systems act autonomously?
What needs to be robust before teams can safely move faster?
These are not abstract questions. They are practical ones that teams are already facing, often without a clear playbook.
This is where in-person community events create real value. They give space to slow down, compare mental models, and stress-test assumptions with people who are working through similar challenges. Not to arrive at definitive answers, but to sharpen judgment and shared understanding.
Thank you to everyone who joined us, contributed openly, and stayed long after the presentations finished. Those conversations that continue beyond the agenda are often where the most useful learning happens.