As reported by Coinpedia, Tulia shared this view during the AI Impact in Engineering panel at Tech Race Summit 2026 in Warsaw. The discussion dealt with a shift that many technology companies are only beginning to face. AI today mostly helps with drafting and analysis, but the next generation of agents will be connected to live systems, including sensitive data and deployment pipelines, and will be able to take actions there directly.

For students and working engineers, the practical consequence is a change in what counts as good work. Tulia expects that by 2029 AI will generate most production code and that smaller teams will manage larger areas of responsibility. When machines produce code quickly, the scarce skills become understanding the business problem and verifying that the output behaves correctly once it reaches production. He suggested that leaders should give teams business context and a clear expected outcome, and then judge productivity by correctness, maintainability, security and operational performance. The number of lines written tells very little about the value delivered.

The same logic applies to the agents themselves. “The more authority we give machines, the more important accountability becomes,” Tulia said. He described a situation that is close to reality for many teams: an agent that can prepare a change and deploy it to production. Two questions follow immediately. Should the agent be allowed to release it without a human approving it, and who is responsible if the deployment fails? In his view, the organisation should answer both before granting access. That means setting permission controls that limit what the agent can do, keeping audit logs of its actions, having a reliable way to stop it and being able to recover quickly from a broken release. The wider industry is having a similar conversation. Anthropic CEO Dario Amodei has called for slowing capability development so that safety work can catch up, while Tulia focused on the limits a single company should set for its own production systems.

Read: Top 5 PWA Development Trends Every E-commerce Brand Should Adopt Now

Preparing for this future starts with foundations that make change safe. Tulia advised CTOs to connect AI spending to a concrete need inside the organisation. His priorities include strong APIs and reliable data, supported by automated testing, observability, security and flexible architecture. Work on architecture and on reducing vendor lock-in may bring little revenue in the short term, but it makes it far easier to switch providers or redesign a system when assumptions prove wrong. He also warned against roadmaps that use up all available capacity, since teams then have no time to experiment with emerging tools or react when priorities move. “I don’t need to predict the future perfectly. I need to make being wrong cheap,” he said.

For the CTO role, the conclusion is that technical depth still matters, combined with a solid grasp of the business. As building technology gets easier, companies will bring in more vendors and more AI-generated systems, and someone has to evaluate them. “I think technical judgment becomes even more important,” Tulia said. Coinspaid Dev, where he works, is an independently owned software engineering company specialising in blockchain.

The post What AI Agents mean for Engineers: Alexey Tulia on Ownership, Safeguards and Technical Judgment appeared first on Reliable Application.

According to Hackread, Tulia, Executive Leader at Coinspaid Dev, joined the AI Impact in Engineering panel to discuss how engineering roles and the job of the CTO are shifting as AI systems gain the ability to act through a company’s infrastructure. His view of the near future starts with money and architecture, then works its way toward accountability.

Why the next budget should buy flexibility

Tulia’s advice to CTOs was to tie every AI investment to a specific organisational need. The areas he prioritised are foundational: well-designed APIs, reliable data, automated testing, observability, security and an architecture that can be reshaped without a rewrite. These are what allow a company to introduce new technology without putting existing operations at risk.

Some of this work, such as reducing dependence on a single vendor, brings in almost no revenue at first. It pays off when a provider has to be changed or a system redesigned because conditions shifted. Tulia summed up the logic in one line: “I don’t need to predict the future perfectly. I need to make being wrong cheap.”

The same thinking applies to people’s time. If a roadmap fills every available hour, the team cannot experiment with an emerging tool or adapt when the business changes direction. Leaving capacity for tests is part of the investment.

How the engineer’s job is changing

AI already speeds up writing code and building prototypes. Tulia sees that speed as an opening for engineers to go deeper into the problem the business is trying to solve and to follow their work all the way into production, where it either delivers or fails.

For this to happen, managers have to supply the missing context: why a task matters and what outcome is expected. Measurement changes too. Output volume says little when an assistant can generate thousands of lines in minutes, so Tulia pointed to correctness, maintainability, security and operational performance as the qualities worth tracking.

When an agent can deploy

Most organisations today use AI to draft and analyse. The next step connects agents to production environments, including sensitive data and deployment pipelines, so they can carry out tasks on their own. The debate around agentic AI has grown sharper this year, with voices inside the industry arguing that capability is moving faster than safety research. Tulia brought that concern down to the level of daily operations.

“The more authority we give machines, the more important accountability becomes,” he said.

He illustrated the problem with a familiar scenario. An agent prepares a change and is technically able to release it. Does it need a person to approve the release? If the deployment breaks a service, whose failure is it? Before such access is granted, Tulia argued, a company should be able to answer yes to four questions:

1. Can we limit exactly what the agent is allowed to do?
2. Is every action it takes recorded in an audit log?
3. Can we stop it immediately if something looks wrong?
4. Can we roll back and recover from a failed deployment?

Without these controls, autonomy simply moves risk around. With them, authority is defined and a human remains answerable for the outcome.

Read: Why Modern Brands Are Moving Away from Basic Warehousing to Value-Added Logistics

A smaller team with a wider remit

Looking toward 2029, Tulia expects engineering teams to shrink while the scope each one covers expands. He also anticipates that AI will produce most of the code running in production. In that world, the scarce skill is checking whether generated work is right.

“I think technical judgment becomes even more important,” Tulia said.

He does not expect the CTO role to become less technical. It will still demand deep engineering expertise, paired with a solid grasp of the business, and it will involve managing a growing number of vendors and AI-built systems as creating technology becomes easier. The practical takeaway for leaders is immediate: decide who owns what, and put safeguards in place, before agents are connected to anything critical.

About Coinspaid Dev

Coinspaid Dev is an independently owned and operated software engineering company specialising in blockchain infrastructure. It employs more than 120 engineers and has over 11 years of industry experience, combining software engineering, infrastructure, security and R&D teams that build distributed systems for more than 20 blockchain networks.

The post Before AI Agents Touch Production: Alexey Tulia’s Case for Cheap Mistakes and Clear Ownership appeared first on Reliable Application.