DrupalCamp Poland 2026: AI Hallucinations, Drupal System Failures and Agentic Maintenance

Three DrupalCamp Poland 2026 sessions on 12 September will examine how systems that appear to be functioning correctly can still produce unreliable outcomes. Piotr Kamieniecki, an independent machine-learning researcher, will speak about AI hallucinations and system-level failures in Drupal, while Paweł Bogucki, CTO at Anri, will examine agentic support and maintenance. The DropTimes spoke with both speakers ahead of DrupalCamp Poland 2026 in Wrocław about the reasoning, human oversight, and engineering judgement behind those topics.

The sessions approach the same underlying problem from different systems. A language model can produce fluent output that is wrong, individually reasonable Drupal decisions can combine into serious performance problems, and automated maintenance can complete more activity without necessarily creating more value. Piotr focuses on building better mental models where a single obvious cause is missing, while Paweł examines how automation should be judged against service quality, organisational knowledge, and the work people still need to perform.

Piotr's first presentation, I Have Become a Language Model, the Confident Idiot. How AI "thinks" and why hallucinations are expected, is designed to build an intuitive understanding of language models without requiring deep machine-learning knowledge. The live conference programme lists the English-language session from 11:45–12:30 CEST in Room 1. Rather than explaining generation only through slides, Piotr plans to make the audience participate in a simulation of how a language model produces text.

The distinction he wants attendees to understand is between using a language model and having a useful mental model of what it is doing. In his explanation, next-token prediction, sampling, and available context help explain why a model can produce language that sounds confident and coherent without guaranteeing that the underlying claim is correct. Hallucination therefore becomes easier to reason about when the model is understood as a system generating plausible continuations rather than as a conventional store of verified knowledge.

Piotr said large language models are becoming part of everyday development work while a substantial gap remains between using them and understanding their behaviour. The interactive format is intended to make that behaviour memorable rather than requiring attendees to absorb a mathematical explanation of sampling. He said he is particularly looking forward to turning a room full of attendees into what he described as “a very inefficient language model,” allowing the mechanics of generation to become something participants experience rather than only hear about.

His second presentation changes the technology but keeps the focus on misleading signals of correctness. When Everything Works Correctly — and the System Still Fails is scheduled from 15:10–15:55 CEST in Room 2 and is based on a technical assessment of a high-traffic Drupal platform. The system developed severe performance and stability problems even though the failures could not be reduced to one obviously defective module or piece of code.

The official session description traces the problem across several parts of the platform. Recurring data imports, synchronous external integrations triggered on content save, aggressive cache invalidation, role-based cache fragmentation, and heavily duplicated Views configurations interacted under traffic and concurrency, contributing to cache stampedes, database contention, and inconsistent experiences for editors and visitors. The important point is not that each of those patterns is inherently wrong, but that their combined behaviour created a system-level problem that could not be understood by examining them only in isolation.

That changes how a performance investigation has to be approached. Looking for a single slow query, adding more caching, or identifying one expensive function may not resolve a failure whose cause exists in the relationships between otherwise reasonable decisions. Piotr's session is therefore aimed at recognising those pressure points early and identifying changes that can improve stability and predictability without assuming that the entire platform has to be rebuilt.

Piotr said he is especially interested in failures that do not produce a neat exception or identify an obviously broken component. Those cases force engineers to reason about behaviour across the system instead of treating debugging as a search for one defective part. For developers and technical leads operating medium- and large-scale Drupal platforms, that distinction can affect where they look first when a site becomes unpredictable under genuine editorial activity and traffic.

Paweł approaches the question of reliable outcomes from another direction in Enhancing support and maintenance services with an agentic layer, scheduled from 13:00–13:45 CEST in Room 1. The English-language session will cover support and maintenance models, agents and human-in-the-loop systems, semi-agentic maintenance, organisational knowledge, budgets, staffing, scalability, and the near future of maintenance services. Its published scope places operational and commercial questions alongside the technical capabilities of AI agents.

The session description distinguishes between automation that performs activity and automation that creates useful results. Paweł plans to use practical examples to examine which applications provide genuine value for agencies and clients and which create what the session description calls an “illusion of productivity.” That framing makes the amount of automation a secondary question to whether an agent actually improves the service being delivered.

Human involvement remains part of that calculation. The session includes human-in-the-loop and semi-agentic models as well as what Paweł describes as the “human premium,” with attention to the balance between AI agents and work where people remain difficult to replace. Budgets, staffing, scalability, organisational knowledge, and service quality therefore become part of the architecture of an agentic maintenance model rather than concerns that sit outside the technology.

Paweł's response to The DropTimes focused on those conversations beyond the formal session itself. He sees the camp as an opportunity to meet Drupal practitioners working in different industries and to hear concerns that may not become part of a prepared conference presentation. That matters for subjects such as agentic maintenance because decisions about automation are shaped not only by what technology can perform, but also by what agencies, engineers, and clients are prepared to trust and take responsibility for.

Piotr also said he is looking forward to meeting people and seeing what problems the Drupal community is working on this year. For him, conversations outside the sessions can be as useful as the presentations because they expose engineers to problems and experiences beyond their own projects. That shared interest in informal discussion gives the technical programme another purpose: comparing how systems behave after they leave controlled demonstrations and encounter real organisations, traffic, workflows, and expectations.

The three sessions do not argue against language models, established Drupal practices, or automation. Their common challenge is to look beyond evidence that something merely appears to be functioning: fluent output still needs to be judged for reliability, architectural patterns still need to be evaluated in combination and under load, and automated maintenance still needs to demonstrate value beyond the number of tasks an agent can perform. For Drupal teams, those decisions ultimately depend on engineering judgement rather than the presence or absence of a particular technology.

DrupalCamp Poland 2026 takes place on 12 September 2026 at Wrocław Technology Park in Wrocław, Poland. Attendees should consult the live conference programme for the latest session times because individual session pages may not yet reflect schedule changes. More information about the programme, speakers, tickets, and venue is available on the official DrupalCamp Poland 2026 website.