Agentic Nesting offers a new approach to enterprise application integration through a multi-agent architecture. Instead of API connections — semantic orchestration through agents.
The problem with traditional enterprise application integration is not the lack of tools, but their extreme rigidity. ESB, API gateway, iPaaS, and RPA address connectivity but increase coupling and the cost of changes. Each new integration requires explicit descriptions of schemas, contracts, and flows. As the number of systems grows, complexity increases non-linearly, and any changes in business logic lead to cascading adjustments. Additionally, such approaches struggle with unstructured tasks and lack an understanding of business context. As a result, systems remain isolated, and users compensate with manual processes.
The proposed Agentic Nesting approach redefines the basic unit of integration. Instead of “Application-as-Tool,” a model of Application-as-Agent is introduced. Each application is wrapped in an AI agent that understands the structure, data, and behavior of the system. Integration occurs not through API calls, but through dialogue between agents — “Conversation-as-Integration.” This reduces dependence on rigid contracts but introduces new trade-offs. The system becomes more flexible and better at handling uncertain tasks, but increases latency and execution costs due to LLM inference.
Architecturally, the solution is implemented as a hierarchical multi-agent system. Unlike “flat” models, a nested structure is used. There are three levels:
- coordinating agent — interprets intent and manages tasks
- application agents — encapsulate specific systems
- capability agents — provide common functions
The key mechanism is Application Agent Initialization (AAI). It automatically extracts the application structure, builds a knowledge graph, and generates a description of the agent with defined capabilities. This is important: the agent receives not just API access, but a semantic model of the system. Interaction between agents occurs through natural language, simplifying system evolution — changes do not require contract reassembly.
At the implementation level, a layered architecture is used. The frontend is built on Next.js with a unified conversational interface. The backend manages the lifecycle of agents and task routing. Each agent runs in an isolated environment, ensuring security isolation and access control. Communication is implemented through NDJSON streaming, allowing real-time tracking of task execution. Support for various runtimes is also provided: CLI agents, API integrations, and protocols like ACP.
Results demonstrate the practical applicability of the approach, but without precise metrics. In use cases, the system automates scenarios that previously required manual coordination between multiple systems. For example, multi-source data analysis or end-to-end workflows are executed through a single natural language request. This reduces operational costs and minimizes errors. However, the authors clearly indicate limitations: increased latency, the complexity of ensuring determinism, and scaling issues with an increasing number of agents.
It is also worth noting the concept of “Application-as-Data.” Connecting an application effectively means connecting its data without ETL. This shifts the architecture from centralized data lakes to a federated access model. Such an approach reduces storage and synchronization costs but shifts complexity to the orchestration and semantic alignment layer.
Compared to the Tool-Use paradigm, Agentic Nesting excels in flexibility and handling complex tasks but falls short in predictability and efficiency for simple operations. Therefore, a hybrid model is feasible in practice: APIs for deterministic tasks, agents for orchestration and reasoning.
In summary: Agentic Nesting represents an evolutionary shift from integration through interfaces to integration through semantics. The approach appears pragmatic for complex enterprise landscapes but requires mature mechanisms for control, security, and cost optimization.
Information source
arXiv is the largest open preprint repository (since 1991, under the auspices of Cornell), where researchers quickly post working versions of papers; the materials are publicly accessible but do not undergo full peer review, so results should be considered preliminary and, where possible, checked against updated versions or peer‑reviewed journals. arxiv.org