Agents everywhere. Developers in danger…

Do you remember how it all started?

I have been working in web development for many years. It all began in the distant 2000s: Denver, Joomla with its templates and plugins. And, of course, nights between raids in MMORPGs, when the first knowledge was being gained.

After a long break, I returned to development and was horrified to discover that my knowledge had irreversibly outdated: it was no longer enough to know how to create websites on popular CMS – now knowledge of frameworks is required. Backend and frontend have become different specialties, and the number of frameworks and popular programming languages is now approaching infinity.

I delved into modern development with fiery interest. My knowledge solidified, but the tasks kept increasing. Finally, I reached a stage where I could work fully and deliver results. I felt a sense of efficiency and expertise! But anxiety remained: competition was growing like yeast — there were just too many programming courses around.

And then came the prime era of artificial intelligence and generative models — ChatGPT appeared.

From disdain to passion

I clearly remember the years 2022–2023. With a slight feeling of disdain, I watched my colleagues, who literally turned to ChatGPT for every little thing. It seemed they didn't want to understand anything themselves — right during the conversation, to keep the topic going, they would go to the chat and generate “their own opinion”. And what about studying, understanding, analyzing — do we really not need that?! I have always valued development for the wonderful feeling of enlightenment after a long search for a solution.

Only two years have passed, and now my best conversational partner has become a chat generator)) AI has replaced search engines, StackOverflow, and almost replaced humans (my dear wife, I hope you are not reading this). I have also mastered an approach that accelerates the process of finding solutions and thinking in general. This approach provides quick access to information that previously took me weeks to search for and absorb. AI has become a wonderful multiplier of my experience. I felt trendy again until I sensed a new race — AI agents.

It seems that I have finally reached a state where I feel like a true specialist. The code of most popular systems no longer seems like magic to me. I seem to have learned to see architecture, understand the consequences of decisions, and confidently review others' pull requests. But a moment of “crystallization of fear” occurred: I am starting to realize that modern development has moved forward without me — now the code is written by steroid-infused agents.

A Grim Picture and a Point of Choice

My information agenda of recent years painted a rather harsh picture — they say that my code is no longer important and can be written by a housewife. Moreover, her younger daughter can design the architecture, and the dog can generate ideas for a startup. About such a grim, somewhat caricatured, but still frightening picture was drawn in my mind.

I understood that there were two options. Either I start to immerse myself in this topic and truly study it, or I remain on the sidelines, trembling and waiting for someone to call me… well, let’s say, to clean the yard or dig graves in the programmers' cemetery.

The choice is to dig into information and delve into the topic.

The First Conference: A Management Perspective

I was lucky: almost immediately, I came across a conference held by Yandex. I attended, was impressed, and even wrote a report article: AI Dev Day — the AI fever continues.

But the conference turned out to be more about management: people who implement AI in their companies, prepare tools for developers, measure effectiveness, and convince businesses of the feasibility of such solutions spoke. For me, it was a useful top-down perspective, but I felt that I lacked practice: “how do I take and start using agents in my daily work?”

Live Sources: Bloggers

At the same time, I found several bloggers who professionally deal with neural networks: they study their capabilities, exploit them, and share real cases. They run channels on Telegram, and the most valuable part for me turned out to be not even the posts themselves, but the comments and chats. There, people discuss nuances that you can't find anywhere else: how to integrate an agent with CI/CD, which providers are cheaper, where there are “pitfalls” with security, and what various models are capable of.

Thus, I gradually began to “collect the AI puzzle.”

Russian Techno: An Insider's Perspective

At the end of March, I attended a conference with a musical name: "Russian Techno," organized by MTS in the Sokolniki Park. And here I looked at my entire gap in knowledge and lag behind the times.

I listened to the presentations and suddenly felt acutely: I have no experience! Not just in development overall, but specifically in using AI agents in real code. When one of the speakers asked the audience: “Who among you uses cloud code?” — I couldn’t raise my hand. First of all, I had a salmon croissant in it, and secondly, using AI agents for everyday tasks in development was terra incognita for me. From the talks, I heard again about the implementation of AI agents for reviewing pull requests. Apparently, this is one of the most mature and widespread cases. But the key thing I noticed: the very approach to designing and creating solutions through agents addresses the problems faced by team leads in development.

Sber's Approach: Agent as an Employee

Especially striking was the approach presented by the speaker from Sber. Their implemented AI agents are analyzed… as full-fledged employees, who have their own KPIs. Their effectiveness is measured in person-hours and financial costs. They undergo a “probation period” in some form.

And you know, I suddenly realized where the myth comes from that all developers will be replaced by AI agents. It grows precisely from here: from this metaphor. When a company starts to consider an agent as an equivalent of a junior or mid-level developer who can review code or write tests, then the next logical step is to think about whether to replace this "employee" with a live person. Many companies, it seems to me, simply have not found another, more adequate way to analyze the effectiveness of agents than to apply the same standards as to Homo sapiens.

However, when you think about it, this approach seems obvious. Conducting development and automation based on real business processes is a pragmatic approach. Well, if as a result of research an agent solves the task faster, cheaper, and with acceptable quality — then it is useful. The task can be assigned to the agent, while a person can be directed to solve others.

The idea of accelerating strategic planning was mentioned. If earlier the horizons were 10 years, then 5, then 3 years, now even corporations are thinking about operational adjustments. The strategy is becoming flexible, and AI agents are not just a tool, but a driver of this flexibility. The approach of "agent as an employee" is an attempt to fit technologies into new compressed management cycles.

At the same time, it is worth noting that the role of the product owner is increasing. Because, despite all the advantages of AI, only a person who truly understands the solution they support will be able to effectively manage the work of the agent. The agent is a tool, but setting tasks, checking results, and taking responsibility for business logic remains with the person.

Safety First

The security issues with the arrival of AI agents have increased, as they can:
- execute dangerous commands,
- substitute artifacts,
- gain access to what they shouldn't.
Therefore, precautions should be stricter, and one should start from a complete ban.

The recommendation that was made from the stage: “use at least rootless Docker (non-superuser mode), but it’s better to move towards a completely isolated infrastructure.” By the way, sandbox services have emerged for running agents in an isolated environment, for example:

• Daytona.

• Modal

• Fly

Here’s a typical scenario for you:
1. set up an environment for a specific task,
2. launch the agent,
3. the agent performs the work,
4. the agent saves artifacts (for example, to cloud storage)
5. the environment is destroyed.

At the same time, secrets (tokens, keys) should be limited to the lifespan of the instance, so that the context for subsequent launches is fresh and secure.

This approach provides an isolated context and maximum security for the working environment. And, it seems to me, this is one of those fundamental principles without which it is impossible to build serious systems with AI agents today.

Who benefits from all this

After conferences, conversations, reading, and reflections, I came to a conclusion that was unexpected for me.

This buzz around AI, vibe coding, and other cursors is beneficial not to businesses and developers, but to the owners of cloud platforms who have kindly prepared wonderful services with per-second billing. This is the infrastructure race of IT giants for users. Undoubtedly, both businesses and developers will find and effectively use this good, but the current conferences are primarily advertisements for cloud services.

It seems that a new reality has arrived where code no longer holds the significance that some programmers attribute to it. What matters is the result of the code's work — the product. Artificial intelligence is the optimization of production costs.

But this optimization has a downside. The deeper a company integrates AI into key processes, the more it becomes dependent on the cloud provider. It reduced engineers, transferred tasks to agents, managers got used to it — and then the supplier raises the price. There’s no turning back now. These are forced costs, not a choice.

Therefore, my prediction is that today's noise benefits cloud giants not only through direct sales but also through the future opportunity to dictate prices. First, make AI indispensable, then—expensive.

Moving forward

One thing is clear: the topic is actively developing. Many teams have already implemented AI agents into their processes and are successfully working with them.

As a developer, this is vital for me. So, I need to experiment, create my own tools, and establish processes. But at the same time, I need to “keep a cool head”—I will still have to work. It seems that a rise in the costs of neural network provider services is not far off)

Understanding how to develop agent systems, and most importantly—why—is where the gold mine lies!

What do you think?