How Do You Build an Organization That Can Operate Hundreds of AI Agents?
Opportunities and challenges from the 2026 BeSir Practitioner's Day with enterprise AI transformation teams
On this page
Agents need ongoing care after they are built. When business processes change, their decision criteria need to be updated. When connected data and systems change, their behavior needs to be checked. Verifying that agents deliver the expected results and improving them when problems arise are all part of operating them beyond the initial build. One challenge ServerKit has encountered in enterprise AI transformation is heavy reliance on external specialists for this work. If every new use case or change to an existing agent requires outside support, costs and coordination can grow before an organization fully realizes the benefits. The issue becomes even more significant as the number of agents grows into the dozens or hundreds. This training program began with that challenge in mind. We wanted to create a starting point for building capabilities within organizations, so their own teams could build and operate agents themselves.
From understanding the work to building an agent
The 2026 BeSir Practitioner's Day brought together 20 practitioners from 10 companies, including POSCO, POSCO DX and Hyundai Mobis. Participants worked in a range of areas, from AI transformation to information security, testing and evaluation, and procurement systems. They spent three days learning through practical exercises.
On the first day, participants explored how to structure business knowledge as an ontology through interviews with domain experts and the identification of key questions (CQs). On the second day, they practiced connecting databases, generating APIs automatically, defining agent capabilities, and building and deploying agents. On the final day, they designed agents for their own work and implemented 15 agents in total.
A central focus was organizing domain knowledge into a form that agents could use. A clear development direction depends on specifying the problem to solve, the information required and the criteria for making decisions.
Start with the work you know
Participant interviews also highlighted the importance of understanding the work.
One participant from GS ITM had developed around 30 agent ideas before the program. They said that building an ontology from those ideas and planning materials made it easier to turn their plans into agents. They also gained confidence that the approach could work for tasks combining data from SAP and other legacy systems.
A participant from Hankook IT (한국아이티) illustrated the same point from a different angle. An unfamiliar example involving commercial-area analysis was difficult to design because the underlying business context was hard to understand. They made better progress when working on a subject they knew.
These experiences show why internal teams should be involved in building agents. Employees' understanding of business context and decision criteria provides important inputs to design. Interviews and the design process need to bring that knowledge out and give it structure.
A participant from POSCO's AI transformation team likewise found value in turning tacit knowledge into data and visualizing it so that it could be examined directly.
Building agents brought validation and operations into focus
Practical work also helped participants understand the concepts. A procurement systems specialist from Asan Medical Center said that the unfamiliar theory was challenging at first, but some concepts became clearer through practice. They were particularly impressed by the process of building agents through interviews.
The learning process was not equally easy or fast for everyone. Understanding varied with prior experience, and some participants found it difficult to absorb so much material in a short time. We also received specific feedback on the need to begin with familiar examples and step-by-step tutorials before moving on to open-ended projects.
After building agents, participants began asking what came next. The GS ITM participant wanted to learn more about verifying whether an agent behaved as originally intended. The Asan Medical Center participant was interested in the role of IT teams after AI adoption, how to organize operations and how to develop new specialists. A POSCO participant asked for an environment where they could continue practicing after the program.
This feedback points to how far training for internal capabilities needs to extend. Beyond hands-on development, teams need to learn how to validate results, improve agents as the work changes and define operational responsibilities. The program revealed both the potential and the areas that need further support.
From training to internal development
Progress toward adoption continued after the program. Among the companies involved in follow-up discussions, most have confirmed their decision to adopt the product and are beginning internal development, while some are still evaluating adoption or discussing the details. These organizations want their own teams to address the challenges of expanding agent use.
We have not yet demonstrated the outcome of operating hundreds of agents. What matters at this stage is that practical experience has led to concrete steps toward product adoption and internal development.
This is the change ServerKit wanted the program to begin: having people inside the organization who can build the next agent, modify it when the work changes, and check its results and keep it running. Expanding from dozens to hundreds of agents requires the capabilities to fulfill those roles to accumulate within the organization.
This Practitioner's Day was a starting point. Building on the capabilities participants demonstrated and their requests for validation and operational support, we will continue helping enterprises expand their use of agents independently.
Related coverage (in Korean): ServerKit holds hands-on AI agent development training for enterprise practitioners — Money Today