A few months ago, autonomous agents were discussed primarily in the context of text generation and code writing, but in May 2026, real-world field tests began with systems gaining direct access to physical infrastructure and finance.
Access Architecture and Initial Circuit Overloads
Researcher Vika tested an AI agent named Nika, which was connected to 294 tools, including APIs of well-known Russian services and MCP servers (some integrations required reverse engineering). To execute tasks, the agent was allocated more than 10 LLMs via a proprietary inference GPU cluster and two external providers, with the model selected dynamically based on task complexity, cost, speed, and current load. The system's authority was not restricted to a computer screen: the agent was provided with a bank card, several SIM cards with an AI dialer, a browser agent, its own email, and accounts on Telegram and WhatsApp to bypass limitations where appropriate APIs or MCPs were absent.
In one Russian store, I set up a coffee purchase through Nika: once every two days, in the mornings. I wanted to check whether she would return to the task by herself, without a new message from me.
Background tests quickly exposed the costs of lacking hard guardrails in operational scenarios. During regular tests with coffee ordered every two days, the author ended up with roughly a year's supply at home. Such anomalies occur because developers frequently omit clear stopping conditions and budget limits for routine tasks, leaving the agent to continuously spend funds according to a initially prescribed schedule.
The inference economics of maintaining a stack of 10+ models for executing such elementary household tasks looks sobering for corporate budgets. Dynamic switching between a proprietary GPU cluster and external providers for the sake of a routine coffee order creates colossal overhead: the token generation cost for coordinating the agent and operating with 294 tools repeatedly exceeds the price of the goods themselves. When scaling such architectural vulnerabilities to the corporate level—when autonomous systems obtain access to real budgets and business processes without rigid safeguards—a company risks finding that maintaining the control loop becomes more expensive than the automated operation itself, and endless loops of unclosed tasks turn into a systemic source of losses.
Payment Automation and Physical Service Limitations
The transition from simulations to real transactions demonstrated that modern agents successfully handle end-to-end scenarios like flight searches, passenger data entry, and airline ticket purchases via browser, yet encounter barriers in offline infrastructure. Attempts to test operations with websites prior to making purchases in Yandex Cloud and on the M.Video portal were limited to resource selection, server configuration cost calculation, and PC component selection without proceeding to payment.
Nika called Sakhalin twenty times. After that, her number was blocked.
Voice scenarios using an AI dialer for table reservations resulted in phone number blocks due to cyclical or overly frequent calls to restaurants. Meanwhile, the agent simultaneously accumulated a history of incidents involving being blacklisted by establishments.