SK Telecom said on the 16th that it will give employee numbers to AI agents. It is an experiment in an organizational system that assigns affiliation, job and authority, and manages them from hiring to departure in a way similar to people. It comes close to a declaration that software will be registered as a subject of personnel management. The company did not disclose how many agents received employee numbers or which departments it applies to.
An AI agent here means a program that takes a goal and divides up and carries out the steps on its own, without a person giving instructions each time. The way several agents pass work back and forth to one another is called A2A (Agent-to-Agent).
Before assigning employee numbers, SK Telecom ran a three-month pilot of an in-house program called the "AX Sandbox" in part of its AI CIC organization. It is an experiment in rebuilding existing work on an AI basis from a blank sheet. The company said a "multi-role" approach, in which one person takes on planning, development and design at once together with multiple agents, was possible, and that the time spent on planning fell and decisions came faster. Figures showing how much it fell were not included in the company's explanation.
Jung Jae-hun, CEO of SK Telecom, presented "AX Innovation 2.0" at the "2026 New Icheon Forum" held in the second week of this month at the SKMS Institute in Icheon, Gyeonggi Province. The company is also preparing a governance framework that sets out which data agents may access and how far their security permissions reach. A plan to tie in-house AI development platforms such as A. Biz, Polaris and Playground into one and link them with major internal systems was announced along with it.
At "AWS Summit New York 2026", held on the 17th (local time) at the Javits Convention Center in New York, Darko Mesaroš, senior developer advocate at AWS, said that even when token costs rise companies do not turn back their use of AI, and that the practice of picking smaller and cheaper models to suit the task is spreading. The assessment of an official at the observability firm Observe, met at the same event, was of a different tone. The mood of a year ago, "if it is AI we buy it", has shifted toward buying only what fits the budget. Observe was acquired by Snowflake in February this year.
At the summit, cases of agent adoption in a range of industries, including investment research, games, new drug development and robotics, were introduced. AWS's coding tool "Kiro" uses a "spec-driven development" approach that settles requirements, design and a task list before producing code. It layered on neuro-symbolic AI that checks logical and mathematical contradictions at the design stage, seeking to reduce the problem of models generating plausible falsehoods.
At "EBSC 2026", a conference held by Younglimwon Soft Lab on the 18th, a point was raised from a different angle. Ho Woong-gi, head of the Future Value Realization Division, said that if work that took three hours by hand is finished in one hour with AI, efficiency has risen but productivity has not. He meant that unless it is decided what to fill the remaining two hours with, the savings pile up nowhere in the organization. He also said that as agents increase, software that runs only through APIs without a screen, so-called "dark software", is appearing.
The stage where agents talk directly to each other is not as near as it seems. Meta put service agents into Facebook and Instagram business pages in March 2025, but these agents talk only with users of its own platforms. David Rothschild, an economist at Microsoft Research, laid out two branches for the shape of the coming market, the closed "agentic walled garden" and the open "agentic web", and said that going the open route requires institutional agreement on identity verification, attribution of responsibility, security and standardization first.
Incorporation into the organization began this month, the sense of cost differs from company to company, and conversation between agents is stalled for lack of rules. Warnings that laying on tools without changing work procedures leaves the state where only time is cut and results stay the same came together from several places.
From an office worker's standpoint, what changes right away is not conspicuous work such as approval lines or meeting minutes. Once it is written into the rules which files the program at the next desk can open and what it can decide, the boundaries of jobs are drawn again as well. When SK Telecom's governance document comes out, the permissions table will be worth looking at more than the organizational chart.
