I was looking for a relaxed, developer-first DevOps event close to Finland, and DevOpsDays Vilnius appeared on my radar. It is a two-day, non-profit conference now in its fourth year, and it was held in Kablys, an interesting Roman-styled building from the Soviet era.
Beyond the traditional sessions, the event leans heavily into interactive formats. The Open Space sessions give participants the floor to suggest topics and drive their own discussions. Then there are the Ignite talks, which are five-minute presentations with auto-advancing slides that force speakers to adapt on the fly and deliver punchy insights.
Naturally, the dominating theme this year was AI. But beyond the technical implementations, the most fascinating discussions centered on our feelings, fears, and the cultural shifts AI is forcing upon the industry.
Agents Are Easy, Infra Is Hard
It seems several companies are now developing autonomous AI agents in the cloud to handle tasks like bug fixing, incident reporting, and root-cause analysis. But while most attendees were highly suspicious of handing over any real control to the AI, there was one team leaning in completely. They mentioned they are already using AI agents extensively in write-mode. A team of two developers spend a substantial amount on unlimited token usage, but they estimated it replacing a full team of developers.
However, the practical reality is that building agents for the surrounding infrastructure is hard. Traditional security concepts such as Role-Based Access Control (RBAC) are strongly again in focus. Robust platform engineering skills are still needed.
The End of Engineering Joy?
Many engineers are genuinely worried about losing the code craft, the deep, focused engineering mindset we actually enjoy. There is a real concern about the psychological shift and brain rot. When you read a technical book or dig into a complex system, you build a mental model. Now, developers often just get the final answer without the underlying context.
Quality is taking a hit because people trust AI outputs without properly reviewing them. A QA professional pointed out that bug reports are now so bloated with AI-generated slop that no one can understand them anymore. The AI has destroyed the fundamental clarity in human-2-human communication.
Open source maintainers are also overwhelmed. Because generating code is cheap, people are just copy-pasting what AI tells them and throwing it at repositories.
Rethinking the PR Process
There was an interesting presentation from an OS maintainer who pointed out that our workflows need to adapt. They proposed moving human discussion to the very start of the PR process. Instead of reviewing the final implementation, we should review and discuss the spec and the design, and block AI from these human-to-human discussions. Once the spec is agreed upon, strong verification automation and guardrails should take care of the code quality at the end.
This aligned perfectly with my own presentation. I talked about establishing shared engineering standards and thorough testing processes to guide AI coding agents effectively in complex CI code development. Moreover, I've been thinking that we should utilize more pair work in the prompting phase. Instead of struggling to review massive amounts of AI-generated code at the end, developers could pair up to review and refine the prompt together before anything is even generated.
Where Should We Run Our Stuff?
It was refreshing to have some non-AI discussions as well. The fundamental struggles of platform engineering remain unchanged. I was highly amused when a presenter mentioned that their teams are still endlessly debating whether or not to use trunk-based development. It sounded very familiar.
There was also a lot of discussion about where to actually run software. Running things on-premises is turning out to be a surprisingly valid choice compared to the big American cloud providers. It can be significantly cheaper, especially for GPU workloads now that many companies are starting to acquire hardware for AI and figuring out how to run it effectively outside the public cloud ecosystems. Alongside discussions about European cloud providers, it seems on-prem is definitely not as dead as the industry thought.
Photos by me
- DevOpsDays venue in (cover) and out
- Robert Dzisevič's presentation for AI PoC (The Three-Week AI PoC That Took Three Months)
- Lina Zubyte's presentation on AI slop (Seeing Through the Noise: Quality Is Imperfect)
Photos by DevOpsDays
- Laura on stage (Developing GitHub Actions at Scale: Architecting for QA and AI agents)
Top comments (0)