Oles Petriv (Reface): "The only option in which there will be no global planetary cabal is hundreds of billions of autonomous crypto agents"

We decided to publish the most interesting part of the conversation separately — in a test version. Enjoy!

About the «valley of disappointment» and the new formation companies

— Olesya, it’s 2026. There is an opinion that generative AI is gradually sliding into the so-called «valley of disappointment» — the peak of inflated expectations has been passed. Do you agree?

— If we measure it by human expectations, we have passed the peak of linear expectations. People always try to predict the future linearly. But processes in AI are exponential. You look for a long time: well, the progression is almost linear, something is growing there little by little. But from the moment when the process seems linear to the moment when the graph becomes almost vertical, a very short interval of time passes.

Linear extrapolation models are now breaking down. The paradigm is changing fundamentally. As for AI agents, on the contrary, we are at the peak of expectations, everyone is talking about them.

— And how is this trend already changing the labor market?

— A new type of company is already emerging. These are businesses that consist of just 1, 2, or 3 people who juggle a million tools, LLM agents, and tools in a super-class way. They don’t narrow their horizons to just one role, «I’m a CTO» or «I’m an engineer.» They just do stuff and make money from it.

What two or three people can physically do today with the help of AI, and what they could do 5 years ago, is heaven and earth. There is almost no hierarchy in such teams, there is a completely different organization of processes.

About AI transformation inside Reface: 3 levels of «eiaifixia»

— Reface employs about 200 people. How do you implement AI within the company? Because most people just buy a corporate subscription to ChatGPT, give it to employees and say: «That’s it, we’re AI-first.»

— It doesn’t work like that. Just giving Copilot or ChatGPT are point solutions. At Reface, we’ve come a long way and realized: «eyification» of a company is not just downloading software. It’s an internal product for employees that should be developed according to all the rules of the product approach: with a convenient UI/UX, security, and its own infrastructure.

We have identified three fundamental levels (layers) of this transformation:

1. Infrastructure level

This is a foundation that needs to be poured with the right concrete. We are inclined to believe that companies should invest time in developing their own internal AI platform on their own infrastructure (on-premise or in their own cloud). You can’t sit on the needle with big guys like Microsoft or Anthropic, because they will do everything for vendor lock.

Its own platform (like OpenRouter analogues, solutions like Bifrost or LightLLM) provides the main thing — observability. You clearly see in one place the control of tokens by employees and teams. You know how much money the company actually burned per month.

2. MCP (Model Context Protocol) Server Level

Once the infrastructure is ready, you wrap all the features and services that people use (Slack, Notion, Amplitude, BigQuery, metrics) into MCP servers. But you do this with access level control based on your internal corporate logic, not just downloading something from Google.

3. Skills level and immersion in the workday

This is where the interaction with people begins. And this level requires the deepest immersion. It is not enough to come to the team as a whole. You need to sit down with a specific person and see what their working day looks like.

For example, a marketer or analyst spends an hour going into Slack, checking for an alert, then going into Amplitude, checking the data, going into BigQuery, getting the User ID, and putting it all together. We turn this routine into an automated skill for AI. But for this to happen, the company must have a separate person (or a whole team) — for example, a Head of AI — who has this full-time job for a year or a half in advance.

«We burn tens of thousands of dollars a month on tokens»

— How much does Reface spend on tokens for internal needs? Is this a large amount?

— The numbers are in the tens of thousands of dollars per month. We have certain stoppers, limits for employees, because strange things have happened before. Someone didn’t set up the author review of a commit correctly, the system got stuck, started running the heaviest Claude model (Opus) non-stop — and voila, $2,000 was burned in one go for nothing.

Professional use of top models is very expensive. But there is a paradox: the cost of tokens in the world will increase precisely because the tokens themselves will become cheaper. This is an unobvious inverse correlation. The cheaper a unit of resource becomes, the more of this resource people consume in total. Due to fierce competition between the USA and China (for example, the effect of the DeepSeek model), token prices are pushed below cost.

About super agents of the future and the threat of a «planetary cabal»

— Where is all this going? What do you see as the next step in the development of AI agents?

— The next stage, a kind of apex at the top of this entire pyramid, is the creation of an adaptive LLM agent for each employee.

It should be deployed on the person’s local computer, connected to all corporate MCP servers and the skills library. Its main purpose is to simply silently observe what the person does, learn from it and gradually automate this particular guy. It will be a kind of superagent: your personal PM, HR and bro in one person, who knows everything about the company.

— It sounds cool, but at the same time it’s scary. If AI becomes so omnipotent, where is the guarantee of safety for humanity?

— If we move towards the concentration of superintelligent systems in one physical place, under one financial or political control, it will lead to the creation of weapons much more terrible than all nuclear weapons combined a hundred times. I really don’t want us to reach AGI (artificial general intelligence) in the form of a conditional closed «Mythos», which the developers will not give to anyone except the FBI, CIA or Trump for security reasons. Or in the form of a giant data center in China with 10 nuclear power plants, which will turn the Internet into a red sink. Both options are super fast.

Each of these agents will not be much smarter than a person. Each will be a narrow specialist, just like you and me. They will have their own personal experience, limited resources and, most importantly, their own crypto wallet. As soon as there is no person in this circle, but there is a machine on which code with access to a crypto wallet is deployed, the agent becomes a completely autonomous digital entity. He himself earns money on his balance through his activities and pays for his tokens and computing power from there.

— So, a hypothetical robot agent will be able to copy itself and deploy it on the network?

— Yes. It can deploy itself anywhere — in your browser or hidden on your phone. I have Hermes Agent running on my phone right now, collecting data.

Will there be LLM agents in such a future who set themselves the goal of destroying humanity? There definitely will be. Right now, there are several thousand offended or embittered people in the world who are sitting and writing code for just such an agent. But when hundreds of billions of decentralized «intelligent entities» stand against one «bad» agent, the scale of potential damage is minimized. Our task in this era of artificial intelligence remains the same — simply to survive.