When enterprises attempt to deploy AI agents into actual business operations, the first wall they hit is often not that the models are not smart enough, but the inability to answer a fundamental question: who is using this agent to access what documents and tools, under what conditions? Seo Jong-hoon, CTO of Skelter Labs, summarizes this dilemma as follows: the bottleneck introduced by AI agents is not model performance, but control.
He provides a specific failure scenario: if each team directly calls different large language models, the company level finds it difficult to track where costs occur, and each team's guardrails differ, making it impossible to maintain consistent control levels across the entire organization. This is not a hypothetical issue; it is an actual permission management challenge faced when deploying AI agents after generative AI has diffused within enterprises.
Seo Jong-hoon's argument rests on a judgment: the past AI market focused on training better foundational models, but in the agent era, open models such as OpenAI and Claude have already been validated. Enterprises truly do not need to build their own stronger models; instead, they need to flexibly and securely manage these ready-made models within their own environments.
To ground this argument, Skelter Labs designed the BELLA platform, with the core idea of running the entire lifecycle of AI agents—creation, deployment, and collaboration—on a unified control foundation. The platform consists of four layers: the Identity Layer identifies the currently logged-in end user and passes this information uninterrupted to the final step of agent execution, i.e., search and tool invocation; the Policy Layer uses the user information passed from the Identity Layer to determine access permissions for specific documents, tools, and agents one by one, blocking unauthorized access at the source; the AI Gateway Layer serves as the sole channel for all large model calls and inter-agent calls, centrally executing cost statistics by user and department, budget caps, model routing, and security guardrails; the Asset Layer manages all AI assets including document indexes, MCP tools, workflows, prompts, and skills, dividing sharing and isolation scopes by enterprise, team, and individual levels.
At the defense level, BELLA sets up four checkpoints: Data and Retrieval-Augmented Generation (RAG) integrity protection prevents document chunks from being tampered with and filters prompt injection and hidden characters during registration; Input/Output Guardrails intercept injection attacks and mask Korean-format personal information, API keys, internal IPs, and domains; Agent Behavior Control mandates human approval based on risk levels and executes code in isolated sandboxes to prevent infinite loops or unauthorized tool calls; Gateway Resource Protection limits token quotas and request frequencies by user and key, detecting abnormal consumption patterns.
Seo Jong-hoon further emphasizes that the platform should not be a layer that blocks change, but one that absorbs change. Models and tools are changing rapidly; if the platform binds itself to a single model or cloud provider, it will become a burden during technological iteration. BELLA is built on Kubernetes, supports on-premises deployment and private clouds, and does not rely on specific models or cloud vendors.