Berlin DataTalksClub Group

This event is sponsored by dltHub

📍Location: Zalando Office BHW
Berlin Hedwig-Wachenheimstraße 7 (BHW)
The main entrance is on the west-side of the building (towards Uber Arena). Participants should register at the Helpdesk, and will be picked up from there.

📝 Agenda:

18:00 – Doors Open Pizza, drinks, and networking
18:30 – Opening Short welcome and announcements from the organizers (5 min)

### 18:35 – Talk #1

Building Blocks of Modern Code Agents: Reusable Skills and Role-Based Subagents

Alexey Grigorev — DataTalk.Club

A practical map of modern code agent types and the two key building blocks that make them reliable in real-world development workflows.
Code agents appear in many forms today: chat-based assistants, cloud agents running CI-like workflows, and IDE or terminal agents embedded directly into development environments. In this talk, we’ll map these main categories, discuss what each is good at, and highlight where they commonly fail.

We’ll then zoom in on two core building blocks used in modern agent tooling:

• Skills / playbooks — reusable, step-by-step workflows (e.g. understand the repo → implement a change → run checks → produce a clean diff)
• Subagents — specialized roles such as planner, implementer, reviewer, and tester that split complex work into focused, reliable steps

Format: 20 min talk + 10 min Q&A

### 19:05 – Talk #2

Guardrailed Agents at Scale: Zalando’s Support Agent for Incident Triage and Stakeholder Q&A

Ivan Potapov, Saugandh Karan — Zalando SE

We’ll share how Zalando built and shipped a specialized internal support agent that helps engineering teams answer stakeholder questions and triage operational alerts — without drowning in context or compromising production safety.

The agent enriches user requests with relevant observability context (metrics, logs, traces, recent deploys), summarizes what matters, and produces a severity assessment with recommended next steps.

A key challenge is context explosion: pulling “all the data” quickly becomes noisy, slow, and risky. We’ll walk through the architecture patterns and guardrails that keep the system production-ready:

• Retrieval and ranking strategies
• Strict tool boundaries
• Policy-driven response formats
• Evaluation checks to prevent overconfident or unsafe guidance

Finally, we’ll cover rollout and migration: introducing the agent alongside existing workflows, aligning it with architecture guidelines, and iterating based on real incident feedback.

Format: 20 min talk + 10 min Q&A

### 19:35 – Pizza, Drinks & Networking

Invite your friends and join our meetup. Special thanks to our hosts Zalando.

​Finally, we'll look at where the field is heading — toward a world where Context Engineering becomes something you train, not just script.
​
​About the Speakers:
​Tobias Lindenbauer is an AI researcher at JetBrains Research, where he advances efficient and effective code agents that robustly solve long-horizon software engineering tasks. Currently, he is most interested in efficiency topics, context management, interpretability and data synthesis. He recently presented “The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management” at the Deep Learning for Code workshop at NeurIPS 2025, highlighting practical pitfalls of LLM summarization-based context strategies and evidence for more computationally efficient alternatives.
​
​Ivan Potapov is a Research Engineer in Discovery Search & Ranking at Zalando, where he builds search retrieval and ranking systems. He teaches data engineering, AI agents, and LLM alignment, with a focus on bridging software engineering and applied ML. His recent work centers on long-running agents and context engineering—memory, state, and retrieval—exploring why many code-first agent designs fall short. His key thesis: context management is becoming something we train and iterate on, not just script. https://blog.ivan.digital/context-engineering-for-agentic-hybrid-applications-why-code-agents-fail-and-how-to-fix-them-076cab699262
​

Join our Slack: https://datatalks.club/slack.html