Who Is Responsible When an AI Agent Breaks Production? Alexey Tulia Weighs In

As covered by TechBullion, Tulia explored this scenario at Tech Race Summit 2026 in Warsaw, speaking on the AI Impact in Engineering panel about the changing roles of engineers and CTOs. His talk fits into a larger public debate about whether AI capabilities are progressing faster than safety measures. The outlet mentions that in September, Anthropic CEO Dario Amodei urged a slowdown in capability development so safety work could catch up. Tulia took the question to the scale of individual businesses, where the practical issue is deciding how far AI agents should be allowed to go inside production systems.

That issue is becoming urgent because the role of AI is changing. For now, companies mostly rely on it to help draft material and analyse data. The next wave will give agents direct access to live systems, including sensitive data and deployment pipelines, so they can act rather than advise. Tulia’s view is that the more power a machine receives, the more carefully accountability has to be arranged. Before an agent is trusted with a production release, the organisation should restrict its permissions, keep detailed audit logs, retain the ability to halt it and have a clear plan for recovering from a failed deployment. More independence for AI in production, he argued, only works when its authority is precisely set and a human stays responsible for what it does.

Tulia also shared a forecast for the rest of the decade. By 2029, he expects smaller engineering teams to oversee larger parts of the business and AI to generate most of the code that runs in production. As a result, the ability to verify that code and to exercise sound technical judgment will gain importance. The CTO will still need deep technical knowledge paired with business understanding, particularly because easier software creation will bring more outside vendors and AI-built systems into every organisation. According to Tulia, technical judgment will count for even more in that environment.

Engineers are already feeling the change. AI shortens the time needed for coding and prototyping, and Tulia sees this as an opening for engineers to learn more about the business problems they solve and to follow their work through to production. Leaders help by explaining the business background of each task and spelling out the result they expect. From there, productivity can be evaluated by correctness, maintainability, security and how well systems perform in operation, instead of by the volume of code a team produces.

Finally, Tulia addressed how technology leaders should spend on AI. He advised linking every investment to a clearly identified organisational need and building the foundations that make change safe: strong APIs, reliable data, automated testing, observability, security and flexible architecture. He also encouraged leaving room in team schedules for experimentation, since a roadmap that uses every resource blocks the testing of new tools and slows the response to shifting priorities. Efforts to avoid vendor lock-in and improve architecture may earn little at first, yet they make replacing a provider or revising a system far easier once assumptions change. His guiding idea is to keep the cost of being wrong low, because perfect prediction is out of reach. Tulia leads at Coinspaid Dev, an independently owned and operated software engineering company specialising in blockchain infrastructure, which employs more than 120 engineers, has over 11 years of industry experience and runs systems across more than 20 blockchain networks.