The New Horizon of "Agentic AI Runtime"
Telco × Agentic AI / Part 6
May 2026 / Author: Hiroyuki Hikino
🎧 Click here to listen to the audio version
If you want to grasp the big picture through audio before reading the main text, please start here.
You can listen just by pressing the play button.
This article represents my personal views and does not represent my affiliated organization.
Introduction: Why I chose the term "Runtime"
By the end of Part 5, it became clear what the industry needs.
To fill the execution layer, we need a single foundation that simultaneously possesses the three elements of brain, hand, and control.
Furthermore, a combination of past approaches (API automation, RPA, standalone LLMs) cannot achieve this.
From here, we finally enter the discussion of a new category.
I call this category "Agentic AI Runtime".
Why "runtime"?
This is not just a matter of terminology choice.
It is a discussion about how to position the new infrastructure layer that the industry will face over the next 10 years.
In this installment, I would like to discuss the meaning of the runtime concept, its architecture, and its positioning within the industry.
The difference between "tools," "software," and "runtimes"
First, I would like to start by clarifying the terminology.
Past approaches have all been discussed within the framework of "tools" or "software".
Ansible is an automation tool
UiPath is RPA software
ChatGPT is an AI tool
However, what fills the execution layer is qualitatively different from these.
That is because it is the "foundation upon which AI agents operate".
This is where an analogy is effective.
The example of Java
Programs once relied on the OS to run.
Java changed that.
By placing a runtime called the "JVM (Java Virtual Machine)" on top of the OS,
programs became able to run anywhere, in the same way, and securely.
The JVM provides memory management, thread management, security boundaries, and exception handling—everything necessary for a program to run. Programs can behave freely on top of the JVM.
Example of containers
The same thing happened when we entered the cloud era.
Container runtimes (Docker, containerd) appeared, and
applications became able to run anywhere, in the same way, and in isolation.
Container runtimes provide networking, storage, CPU allocation, and security boundaries—everything necessary for an application to run.
AI agents also need a runtime
And now, we are entering an era where AI agents "act autonomously."
To support this, we need a runtime that allows AI agents to operate safely, in a controlled manner, perform any operation, and run anywhere.
That is the Agentic AI Runtime.
Definition of Agentic AI Runtime
I would like to clarify the definition here.
Agentic AI Runtime is:
A new infrastructure layer that allows AI agents to execute arbitrary operations, including both APIs and GUIs, in a safely isolated environment that meets enterprise control requirements.
In this paper, within the context of Telco AN, we define the Runtime specifically as the "foundation that establishes the execution layer."
In the subsequent Act 2, we will expand this concept into a cross-industry methodology that includes observation, judgment, execution, learning, and control.
Breaking down the elements, it looks like this.
| Element | Content |
|---|---|
| Foundation for AI agents to run | A layer with the same level of abstraction as an OS / container |
| Isolated execution environment | Starts up ephemerally and does not touch production directly |
| Arbitrary operational capability | Can handle both APIs and GUIs just like a human |
| Enterprise control | Identity, auditing, least privilege, rollback |
This is the category of Agentic AI Runtime.
Runtime architecture: 5 layers
The Agentic AI Runtime consists of 5 layers.
Each layer performs a specific role.
Layer 1: Agent Model (Brain)
The top layer governs the "brain" of the AI agent.
Large Language Models (LLMs)
Computer-Using Agent (CUA) models
Domain-specific fine-tuned models
This is where natural language instructions are understood and procedures are constructed.
Layer 2: Orchestration (Command)
The layer that coordinates multiple agents and multiple operations.
Task decomposition and assignment
Sequence control and parallel execution
Error detection and retry
Escalation to humans
This is the mechanism for agents to work as a team.
Layer 3: Execution Sandbox (Hands)
The place where agents actually "move their hands."
Ephemerally launched Cloud PCs and containers
API clients, GUI operations, and command execution
Retrieval of operation results and feedback to the judgment layer
Here, instructions from the brain are converted into physical operations.
Layer 4: Control and Governance (Governance)
The layer that controls all operations.
Integration with identity infrastructure (Entra ID, IAM, etc.)
Control of least privilege
Automatic generation and storage of audit logs
Policy validation (pre-execution check)
This is the backbone required to withstand production operations.
Layer 5: Observation and Telemetry (Nerves)
The layer that observes the state of the entire runtime.
Telemetry of agent behavior
Visualization of performance, cost, and usage
Detection of anomalous patterns
Data aggregation for auditing and analysis
Here, the runtime itself becomes an "object to be observed."
The runtime only functions when all 5 layers are present
I want to emphasize this here:
an Agentic AI runtime only makes sense when all 5 layers are present.
Past approaches only possessed one or two of these layers.
API automation had Layer 3 (part of the hand) and Layer 4 (control) but was weak in Layer 1 (the head)
RPA had Layer 3 (part of the hand) but was weak in Layers 1, 4, and 5
LLMs alone had Layer 1 (the head) but lacked Layers 2, 3, 4, and 5
A foundation integrating all 5 layers did not exist in the industry.
That is why the execution layer remained a void.
The changes that an Agentic AI runtime will bring to the industry
How will the industry landscape change once this runtime becomes widespread?
Three major changes are expected.
Change 1: Redistribution of Operational Man-hours
Areas that were previously "manual work" will now be executed on the Agentic AI Runtime.
Disaster recovery operations
Configuration changes and setting adjustments
Operations in multi-vendor environments
Routine audits
The human role shifts from "performing tasks", to "designing and supervising agents".
Change 2: Relativization of Vendor Lock-in
Until now, operations were tied to vendor-specific consoles. Since agents on the runtime can operate GUIs just like humans, they are not bound by specific vendor APIs.
This has the potential to change the vendor power balance in the operations industry.
Change 3: Standardization of the "Execution Layer"
The execution layers that were previously fragmented will be standardized through a common abstraction known as the runtime.
Just as the JVM emerged and standardized programming. Just as containers emerged and standardized deployment. With the emergence of the Agentic AI Runtime,
autonomous operations will be standardized.
The industry race for the runtime has already begun
I am not the first to propose this category of runtime.
Major global players are each trying to reach an "Agentic AI Runtime" in their own way. The next 2 to 3 years will be a
battle for supremacy in Agentic AI Runtime.
Hyperscalers
Existing operations infrastructure vendors
AI startups
And, Microsoft
Which of the five layers does each company hold, and where are they trying to reinforce?
I would like to discuss this mapping in detail from the seventh installment onwards.
Why does this have special significance for the telecommunications industry?
The category of runtime itself is a cross-industry concept.
Finance, manufacturing, healthcare — it can be applied to any industry.
However, the telecommunications industry hasunique characteristics not found in other industries.
1. 24/365 operation is a prerequisite
Telecommunications networks cannot be stopped.
Agent execution must continue to run during nights, holidays, and exceptional situations.
2. Extreme multi-vendor mixed environments
RAN, core, transport, OSS — it is never provided by a single company.
The runtime requires extremely highheterogeneous environment support capabilities.
3. Strict regulatory and audit requirements
The telecommunications industry is under the supervision of regulatory authorities in each country.
The control layer of the runtime mustcomply with telecommunications industry standards.
4. The industry goal of reaching L4
There is a clearer goal than in other industries (TM Forum AN L4), and the entire industry is moving in unison.
The runtime is asubstantial prerequisitefor reaching L4.
These unique characteristics make the telecommunications industrythe most urgent application area for Agentic AI runtime.
Conclusion: From here on, we will finally step into the "concrete"
Over the past six installments, I have organized the composition of the industry.
The wall of autonomous operation (Part 1)
Conditions for reaching AN Level 4 (Part 2)
The void of the execution layer (Part 3)
Three elements required for the execution layer (Part 4)
The inevitability of why existing approaches did not align (Part 5)
The New Horizon of Agentic AI Runtime (Part 6)
This concludes Phase 1 — raising issues as a spokesperson for the industry.
From next time, we will enter Phase 2.
We will discuss specific players, specific technologies, and specific solution approaches.
The "foundation to fill the execution layer" that the industry has long awaited is no longer just a concept.
It is beginning to emerge as a real product.
I would like to observe its form together with everyone in the industry.
Next time: "The Emergence of a Foundation Supporting the Execution Layer — The True Identity of What the Industry Has Been Waiting For"
This article: approx. 4,900 characters / 14-minute read
Author: Hiroyuki Hikino
20 years in the telecommunications industry. 3G/4G/5G mobile core network design / Microsoft (personal views) / Representative of Diestaday LLC
Please feel free to leave your opinions and questions in the article comments.