DevOpsDays 2026

Vilnius, Lithuania

30 Sep 2026

To be completely honest, I’ve been feeling overwhelmed with everything happening in the world these days (AI, layoffs, the wars), and arranging a trip to a conference abroad seemed like too much. But I still wanted to stick to the idea of visiting one conference a year. So I attended DevOpsDays in Vilnius. I appreciate that it is not a profit-oriented event and any leftover profits are donated.

This was the 4th time the conference was organised. It lasted 2 days. Each day consisted of 5 x 40-minute talks, 4 x 5-minute ignite sessions, and open space discussions (which I skipped).

Reflections on the talks

I won’t discuss every single talk here, just the ones that stood out to me the most.

Day 1

Jan Moser – How we build an airport with AI. One genre of videos that I like watching on YouTube is megaprojects - various architectural ideas that are innovative either in their construction or in their social implications (channels like TheB1M or Megaprojects). This talk would fit nicely into one of those. The use of AI is two-fold here: in the daily operations of the airport as well as in producing documents for building it. The daily operations part is rather scary. The airport is being built in Vietnam, where the privacy policies are lax, and there will be a lot of biometric tracking and human recognition algorithms applied. The system will supposedly be able to recognise a person in a crowd based only on their body posture!

The document part was a lot more mundane. AI was used to generate some 700 000 lines of markdown that were eventually turned into construction specifications. I assume they worked on this for a while, through various stages of model quality. The lessons learned seemed quite simple, for example:

One really good reminder was to not default to using LLMs for everything. Sometimes it is much cheaper and more practical to use an LLM to help you write a script that you can then simply re-run and get a good, deterministic output. Let’s not forget our developer skillset and mindset!

Grzegorz Kalwig – Wrong tool, Wrong Quest: What IT Teams Can Learn from RPGs. I feel that with the current AI image generation capabilities, every conference now has at least one talk where people try to merge software engineering practices with fantasy/RPG themes. I’m not a big fan of this approach, as the rich visual theme may drown out the actual technical message. But one thing I liked in this talk was the stats for various aspects of systems (operational excellence, security, reliability, performance efficiency, cost optimisation, sustainability) - the six pillars of the AWS Well-Architected Framework. I think the levels of excellence were captured well in these. A few examples:

Points Operational Excellence Reliability Security
1 Manual deployments, no structured logging Single-AZ deployment, no failover Basic IAM roles, default settings
2 Basic CloudWatch alarms, semi-automated scripts Multi-AZ, simple retry logic Separation of environments, WAF
3 IaC (SAM/CDK), versioned deployments, basic CI/CD Failover strategy + DLQ, backup & restore in place Fine-grained IAM, Secrets Manager with automated secrets rollout
4 Full CI/CD pipelines with rollback, metric-based alarms Cross-region backup, chaos testing Policy boundaries, Shield Adv., MFA enforced
5 GitOps/IaC + self-healing + automated chaos experiments Warm standby multi-region architecture Zero trust, automated compliance checks (Config/SCPs), audit logging

The argument was that, as in RPG games, you are not allowed to set every stat to the maximum value. In real engineering life, time and money stop you from doing the same with your system.

Saulius Valatka – V-Assist: AI agents for all at Vinted. I work at Vinted and use V-Assist daily for almost every single thing I do. I’m a huge fan of it. But what I appreciated most in this talk was Saulius’ honesty about the current state of working with AI. The gist of it was that while individual developers may feel like 10x developers with AI, the pace of delivering big initiatives is mostly unchanged. That’s because the structure of the work is still the same. We attend meetings, agree on what to do and refine the implementation details before passing them on to an agent. Then we wait for the agent, usually watching YouTube in the meantime, and give feedback. Rinse and repeat. At best we might run several tasks like this in parallel, but then we pay the huge price of context switching. Humans are generally happiest when they are able to do deep work. So ideally we would have one focused design session, hand the result to AI and move on to other things without needing to check in constantly.

Konstantinas Jurgilas – So You Want to Replace a Platform: Lessons Learned from Inside the Trenches. I want as many talks like this as possible. People making a big, long-lasting change to widely adopted internal tooling and then sharing their reflections afterwards. I think it is very valuable, as this is something platform engineers have to do often, so the more we can learn from others, the better.

The first part of the talk was dedicated to the ‘people’ side of things. I loved the idea of a ‘resistance budget’. Similar to how errors eat at your SLO budget, bugs and missing feature parity will eat at the resistance budget (maybe it should be called a patience budget?). Improvements and well-provided guidance, on the other hand, will increase it. The suggested forms of guidance were:

For the technical part, the emphasis was on automation. The automation can be done with AI skills and MCP servers. I’ve been a part of several migrations where a bunch of teams had to adapt to new tooling. Often we provided documentation and asked them to implement and verify. It could have been so much simpler with AI tools! We could have easily written AI skills to do the code changes, since they usually come down to very specific steps like replacing X with Y and renaming some ENV vars. AI could also apply these changes to every repository and open pull requests for the teams. Given a specific checklist, it could even monitor Grafana dashboards for anything out of the ordinary. An important point in this talk was that if we automate the execution, we should also automate the verification - otherwise verification becomes the bottleneck.

Day 2

I noticed a tendency with events that last multiple days. During the first day I absorb a lot more information, but I also get overstimulated and don’t sleep as well afterwards. On the second day I’m really tired and find it much harder to enjoy the event. It’s the same energy drain I wrote about after GoLab. I think this is part of the reason only one talk really stuck with me on Day 2.

Josephine Pfeiffer – Touching Grass (and 40-year-old C code). Josephine is one of the maintainers of the NetworkManager project. She spoke about the pains that AI has caused for the maintainers of open source projects. Before AI, comprehending and writing code was the bottleneck. People took a while to make a contribution. Now, they can copy-paste an issue description to Claude and have a pull request ready in no time. Or a slop grenade, as Josephine put it. It seems the AI-generated pull requests are big, and the communication is poor - even the replies are written by AI. It has become so problematic that she wrote a policy for the use of AI agents that included the tenets “Write your own commit messages and Merge Request descriptions. Respond to review comments yourself.” To make it easier to catch the people who do not comply, she also added an instruction to AGENTS.md to add the word biblioklept whenever AI is the one generating the description. This got her a bit of attention on Phoronix news. Mostly this was a fun story to hear. But I also see merit in keeping the communication layer fully human.

Reflections on the ignites

The ignites were speed talks where the presenters had to prepare 20 slides that would flip automatically over the span of 5 minutes. Most of these were nice quick ideas that I don’t have much to reflect on. Except for Lina – Seeing Through the Noise: Quality Is Imperfect. Within those 5 minutes she managed to beautifully capture and express the frustration of being served a lot of slop as a human being. It drove her to retreat to nature and meditation. I can totally relate to that. I suppose slop hits QAs extremely hard, as they get AI-generated features and bug reports. But I guess we have all been in situations where a counterpart in a discussion just copy-pastes an AI-generated response. This is how pages like no-slop.ai are born. I’m a bit upset that the agenda does not include Lina’s full name. I would like to let her know how great an impression the talk left on me. And maybe get hold of the poem she ended the talk with.

Resources

Final thoughts

Overall, in this AI-themed conference, the talks that resonated with me the most were the ones that had a strong human emotion. The frustration of being at the receiving end of AI slop, the honesty about watching YouTube while the agent does the heavy lifting.

As for the conference itself, it is actually a gem that more foreign DevOps engineers should discover. For around 100 EUR you get two days of proper talks (with a lot of international speakers) in a unique venue - Kablys. There were warm meals for lunch, and if you wanted to, you could start buying beers from 2 P.M. (the evening beers were free of charge).

© 2026. All rights reserved.