How we use agents to 10x our data team

We are 3 in the data team at Photoroom, serving 140 colleagues, ~50 data sources and billions of tracking events a month. Most mornings, the night's data incidents are already investigated and fixed in a pull request that another agent has reviewed, before we open our laptops.
Here is the counterintuitive part: the agents are the easy bit. They multiply whatever they are plugged into, so a shaky platform gets a multiplied mess. Our leverage came from making the platform boringly reliable first.
This talk walks through the order we did it in:
1. Opinionated design: one aligned way to work, so automation is safe to run.
2. Tests everywhere: contracts on pre-push, checks in CI, anomaly and semantic-drift detection in prod.
3. Agentic operations: overnight investigation, fixes, and review.

Then the payoff, which reaches beyond our team. Reliable foundations let us hand agents to everyone else: devs ship tracking changes through an agent that validates the plan and opens the Amplitude PR for them, and colleagues get answers by talking to an LLM instead of learning a BI tool. Our Insights Factory runs 100+ query investigations, and self-served analyses turn straight into client-ready emails. We will also show where the agents got it wrong and the guardrails that caught them.
The 3 of us moved up: the agents execute, and we design the system and review what matters. Takeaway: your mandate as a data team is reliability. Build the foundation, then let AI raise the quality of your data and empower everyone around you.

(depending on the need, the content can be intermediate or advanced, but I think it is more inspirational and interesting if we keep it somewhat intermediate)