Agentic AI in the Wild: What Actually Runs in Production
Agentic AI in the Wild: What Actually Runs in Production Connected Intelligence: AI Builders Meetup - WeAreDevelopers PreDay, Berlin
Date: Tuesday, July 8 Time: 17:00 - 20:00 GMT+2 Location: w3.hub Berlin, Berlin, Germany
About Event
AI agents are shipping to production. The question is no longer whether to use them. It's what it actually takes to make them work once the demo is over.
This WeAreDevelopers PreDay meetup is for builders, founders, and practitioners who are deep in the work: wiring AI agents into real codebases, CI/CD pipelines, developer platforms, and knowledge systems. We're gathering the night before the main event to have the honest conversations that don't fit in a conference talk. What breaks at scale. What context engineering actually looks like. Where the failure modes hide.
The theme is connected intelligence: how graph-based reasoning, smarter pipelines, and agentic code review are changing what production AI systems can do, and what engineering teams need to do differently to support them.
And yes, that includes Java. If you've been treating AI as a black-box REST endpoint bolted onto your stack, Ana Maria Mihalceanu's session will change how you think about that. With JDK 25 and the Foreign Function and Memory API, you can own the full inference stack from Java, wire real models directly to native runtimes, and use Project Babylon's code reflection to author pure Java AI-ready models. No glue language. No external model files. Just Java, all the way down.
No vendor pitches. Just the stuff that runs in production.
Agenda
17:00 - Doors Open, Drinks and Networking
17:30 - Welcome and Opening Remarks
17:45 - Tech Talk 1: Neo4j
18:10 - Break, Food (Pizza) and Networking
18:35 - Panel Discussion: Agentic AI in the Wild
19:20 - Tech Talk 2: CircleCI
19:45 - Tech Talk 3: Ana Maria Mihalceanu, Java Champion and Developer Advocate, Docker
20:15 - Closing Remarks
Speakers:
Andreas Kollegger, Director of GenAI, Neo4j
Title: Where is the Zebra? Agent Decision-making As Constraint Satisfaction
Zebra Puzzles are a great exercise for agentic reasoning and decision-making. Solving them requires no domain knowledge — just pure constraint satisfaction. Real world examples, though, are much messier. For example, whether or not to approve a loan application has a combination of hard and soft rules, and maybe allowance for loan officer discretion. What’s an agent to do in the real world? We’ll walk through Zebra Puzzle variants to examine how different architectures approach structured decision-making, and why recognizing constraint satisfaction problems changes how you build agents. You’ll learn: – How constraint networks, LLMs, and hybrid approaches solve the same logical puzzle differently – Why enterprise decisions (resource allocation, approval routing, compliance) hide Zebra Puzzles beneath complexity – Patterns for detecting constraint satisfaction in disguise and choosing between symbolic, neural, and hybrid architectures – When your agent needs search vs. inference vs. generation Real enterprise decisions often contain hidden constraint satisfaction problems. Recognizing the puzzle structure determines whether you need symbolic precision, neural flexibility, or both. You’ll leave knowing how to spot these patterns and build agents that reason systematically under constraints.
Ana Maria Mihalceanu - Java Champion and Developer Advocate, Docker
Ana Maria is a Java Champion, Developer Advocate, and co-founder of the Bucharest Software Craftsmanship Community. She specializes in Java, Kafka, and cloud platforms, contributes to open source projects including JHipster and LangChain4j, and brings 15+ years of consulting experience across telecommunications, banking, and the public sector. An active technical speaker and author.
Talk: Now and Next Java for AI
Tired of treating AI like a black-box REST endpoint? What if you could own the stack: shape the tensors, steer memory, pick execution providers?
With JDK 25, you can wire real models, LLMs, image classifiers, or object detection algorithms, straight from Java via the Foreign Function and Memory API to call native runtimes like ONNX for fast CPU/GPU inference. You will learn how to map tensor buffers to Java MemorySegment, how to flip execution providers, and build a self-contained Java application. Then we push further with Project Babylon's code reflection: express model logic as Java code that Babylon can analyze and lower to accelerator backends, skipping external model files and the need for a glue language.
Build expressive and testable FFM-based inference today and author pure Java AI-ready models with code reflection tomorrow.
Charles Francoise, Staff Software Engineer, CircleCI
Talk: Plan with Opus, Code with Sonnet
What if the key to working smarter with AI agents isn't just the model - it's which model you use for which job?
A simple experiment with Claude Code—planning with Opus, then handing that plan to Sonnet for implementation—turned up stunning results. Is it faster? Is it cheaper? Is it worth it? We'll live-code the experiment together, exploring the variables that make multi-model orchestration work, and we'll demonstrate a repeatable workflow that any AI-forward developer can steal.
Panel: Agentic AI in the Wild
What does it take to get AI agents out of the proof of concept and into a system that runs reliably? This panel brings together practitioners from across the stack to talk about the real constraints: context management, failure modes, governance, and what changes when you scale. Expect specific lessons, not talking points.
Moderator: Dana Fine - Open Source and Community Manager, Qodo
Dana leads open source programs and community at Qodo. She runs the GitHub User Group IL and CNCF Israel communities, organizes the Bond AI meetup series, and has been building developer communities across the cloud native and open source ecosystem for years.
Nnenna Ndukwe - Developer Relations Lead, Qodo
Nnenna leads Developer Relations at Qodo, the AI code review platform. A software developer, applied AI researcher, and community builder with over a decade of engineering experience across med-tech, fintech, and media-tech. A 2019 Google Women Techmakers Scholar, she specializes in integrating AI code review into modern development workflows for open source and enterprise teams at scale.
Sebastian Kister
Sebastian is a cloud pioneer and enterprise transformation practitioner recognized as one of the first to implement a production-ready architecture for Agentic AI Operations in a large enterprise environment. An active CNCF and Linux Foundation member, he advocates for scalable platform ecosystems and a people-first philosophy: people first, then tools, then processes.
Tevfik Aloglu
Tevfik Aloglu works at OpenAI in Applied AI team, partnering with ambitious startups across Germany and Europe to turn frontier AI into product advantage and real business impact. Previously, he co-founded the generative-AI startup Pyne as CPTO and built AI and machine-learning ventures at BCG Digital Ventures and Project A. Tevfik studied computer science at the Technical University of Munich, spent time at Carnegie Mellon University, and is an alumnus of CDTM.
Location
W3.hub Berlin Straßburger Str. 55, 10405 Berlin, Germany