The Execution Layer of Autonomous Networks, and Beyond — Act 1 Summary and Connection to Act 2

Telco × Agentic AI / Part 17 (Act 1 Finale)
August 2026 / By: Hiroyuki Hikino
🎧 Click here to listen to the audio version
If you want to grasp the overall picture through audio before reading the text, please start here.
You can listen just by pressing the play button.

Introduction: Closing Act 1 Here

From the first installment to this piece, this has been a 17-part series.
Including the extra editions, it totals about 70,000 characters. If turned into a book, it would be one slim volume.

In this final installment, I would like to start by compressing the individual discussions I have written so far into a single structural diagram.
Then, I want to show why this series does not end here, but moves toward a separate series called Act 2, and identify the connection point.

As I wrote in Part 12, this series has passed through the most challenging installment along the way.
I would like to express my deepest gratitude once again to the readers who have followed along this far.

Compressing the Act 1 Discussion into a Single Structural Diagram

If we rearrange the entire series in the form of industry problems and solutions, it looks like this.

Problems Facing the Industry (Parts 1–5)

1. Autonomous networks are stuck at L3 (Part 1)

2. The reason is that observation and judgment are within range, but the execution layer is a blank space (Parts 2–3)

3. To fill the execution layer, the three elements of head, hand, and control are needed simultaneously (Part 4)

4. Past approaches (API automation, RPA, standalone LLMs) all possess only two of the three elements (Part 5)

Structure of the Solution (Parts 6–8)

1. As a single foundation possessing all three elements simultaneously, a new category called Agentic AI Runtime is necessary (Part 6)

2. Its implementation is officially coming to the forefront at Build 2026 as Windows 365 for Agents (Part 7)

3. The core of control is being fought on a different main battlefield called Entra Agent ID (Part 7.5)

4. The orchestration layer is appearing as Project Opal (Part 8)

Application to Business (Parts 9–11)

1. Fault response: A closed-loop where MTTR is reduced by an order of magnitude (Part 9)

2. Service Deployment: Overcoming the Vendor GUI Barrier with CUA + Opal (Part 10)

3. Self-Healing/Optimization: The Learning Loop (Especially Procedural Memory) Establishes L4 (Part 11)

Structural Theory and Supplementary Notes (Parts 12–16)

1. Vendors Capable of Meeting the Five Requirements for the Execution Layer Are Structurally Limited (Part 12)

2. The TM Forum AN Framework Is an Organizational Maturity Model, While This Series Focuses on Technical Architecture; They Are Complementary (Part 13)

3. Governance Is Not a Constraint, but the Starting Point of Design (Part 14)

4. The Target of Observation Must Be Elevated from Substrate to Outcome (Part 15)

5. The Detection Layer and Structural Defense Layer Are Separate Tasks; Zero Trust Design Is Only Achieved When Both Are Present (Part 16)

The central axis connecting these is the single sentence: “Filling the void of the execution layer with an Agentic AI Runtime.”

The message of Act 1 in a single sentence

If I were to compress the 17 discussions into one sentence, I would write it like this.

This is the single sentence I wish to present to the industry as Act 1 of this series.

Beyond Act 1 — This is not a topic unique to AN

Here, I would like to elevate the structure of this series one level higher.

This series has used the heading “Telco × Agentic AI.” However, upon re-reading, you will notice that the content of the discussion is divided into
parts unique to Telco and parts that are not.

The right side—the discussion not limited to Telco—is actually a methodology that holds true across industries, merely using Telco as the subject matter. Manufacturing production management, financial transaction monitoring, medical business operations, government counter services, software development operations, and consulting delivery—all possess the structural void of an execution layer.

In other words, Act 1 could be said to be the work of establishing the methodology of an Agentic AI Runtime using the most difficult subject matter: Telco AN.

Connection point to Act 2

From here, this series moves toward the next series — Act 2.

The theme of Act 2 is to deploy the methodology of the Agentic AI Runtime itself across cross-industry layers.

With the emergence of agents running on local devices, such as Microsoft Scout (a desktop AI app) around Build 2026, this connection point has become even clearer.
In Act 1, I discussed the "execution layer on Cloud PC" with a focus on W365A, but in Act 2, it is necessary to broaden the perspective to include agents on local Windows, agents running on Microsoft 365, and the control boundaries that span them.
The execution layer is beginning to appear not as a single product name, but as a Runtime structure that spans cloud, local, and control boundaries.

As a preview, I will list the pillars that will be covered in Act 2.

1. 3+2 Architecture

The basic structure of the Runtime, consisting of 3 Layers (Observation, Decision, Execution) × 2 Loops (Learning / Governance).
While I touched on this partially in the Telco context in Act 1, I will organize it as a cross-industry methodology in Act 2.

Here, I would like to clarify the relationship between the "Head, Hands, and Control" used in Act 1 and the 3+2 Architecture covered in Act 2.

In other words, the "Head, Hands, and Control" of Act 1 was a requirement organization to explain the execution layer of Telco AN, while the 3+2 Architecture of Act 2 is an organization that expands this into a cross-industry Runtime methodology.

2. Agentic Closed Loop Maturity Model (L0 to L6)

A maturity model from the technical side that measures how much an organization has adopted Agentic AI.
I will also re-organize the correspondence with TM Forum AN here as a cross-industry version.

Note that while the L4 of TM Forum AN and the L4 of the ACLMM covered in Act 2 share the same "L4" number, the evaluation targets are different.
The former is the maturity of autonomous operations in the telecommunications industry, while the latter is the stage measuring how much the Agentic AI Runtime has been technically incorporated.

3. 4-Layer Memory (Episodic / Semantic / Procedural / Constitutional)

I will delve into the contents of the Learning Loop by dividing it into four layers.
I will expand on the content that was only mentioned by name in the 11th installment as a cross-industry design guideline.

4. Independence of the Governance Loop

Control is not a layer subordinate to any of observation, decision, or execution.
A design principle that it must be designed as an independent loop that penetrates all layers.
I will extract the discussion from the 14th installment and rewrite it as a general theory of enterprise design, moving away from the Telco context.

5. Human Role Shift (in-the-loop → on-the-loop → above-the-loop)

The position of human intervention rises from task execution to purpose, constraint, and control design.
This is a discussion that redefines the "relationship between humans and AI" across industries.

6. Discussion on the deployability of AGI

I will place the sharpest argument of Act 2 here.

This does not negate the current AGI debate in the industry (model scale and scaling laws).
Rather, apart from the debate on how capabilities emerge, this is a structural response from the operational side regarding how to make them deployable within society, enterprises, and regulations.
The argument established in Act 1 of this series—that the 'execution layer is a void'—emerges here as a question from the field regarding the AGI debate.

Just one request to the industry

Finally, I would like to make one request to those in the industry before closing Act 1.

The greatest purpose of this series is for the terms discussed—'execution layer,' 'Agentic AI Runtime,' 'three elements,' 'head, hands, and control,' 'four-layer memory,' and 'Governance Loop'—to become used in industry discussions.

Whether you agree, disagree, criticize, or offer improvements, it does not matter.
As a common language for discussion, having these terms live within the industry will be the first and greatest contribution toward the entire industry reaching L4.

It generally takes one to two years for industry terminology to become established.
If this series can serve as the starting point for that first year and assist in the discussions of those in the industry, there would be no greater joy.

In conclusion

I will close the 17-part Act 1 here.

Autonomous operations in the telecommunications industry will become the site of social implementation for Agentic AI Runtime sooner than the AGI debate—this is the message I most strongly wanted to convey throughout this series.

Telco is the domain with the strictest infrastructure requirements, the highest control requirements, and the heaviest social responsibilities in the industry.
If an Agentic AI Runtime can be established there, it can be deployed far more easily in other industries.
Conversely, what cannot be established in Telco cannot be fully operated in any other industry.

Therefore, I want to make this happen together with those in the industry.

In Act 2, I will broaden the perspective to cross-industry and continue the same discussion.

Next time preview — but before that, I will insert an 'intermission'

However, before entering Act 2 immediately, I would like to take a breath just once.

The next installment (Part 17.5) will be an intermission where I will not talk about telecommunications at all.
The subject will be my own body.

For the past four months, immediately after joining Microsoft, I have been running another loop in parallel with writing this series.
90.8kg on April 6, 2026. 74.6kg on August 3, 2026.▲16.2kg.

I only did two things: take a photo of every meal and weigh myself every morning.
Other than that, I was just running a loop with AI based on the data generated from that.

However, looking back, what was happening there was word-for-word the same as the structure I spent 70,000 characters discussing in Act 1.

• There was a void in the execution layer where 'I knew it, but I couldn't do it.'

• Observations that cannot be sustained are as good as non-existent (The Wall of Observation)

• Left to their own devices, decisions naturally split into three layers (Rules/LLM/Human)

• Repeated decisions eventually crystallized into rules

• Over-applying them nearly broke the body itself (Design of Control)

• And the correct answer itself began to drift along the way

Why write this in a technical series?

In this article, I wrote that 'the discussion in Act 1 is not unique to Telco.'
This is because I want to prove that claim using a subject that does not use a single piece of technical jargon.

Autonomous operational mechanisms are not a specialized technology of the telecommunications industry.
The same form holds true in manufacturing, healthcare, distribution, retail, and back-office operations.
Telecommunications just happens to be the field where these problems appeared earliest and most intensely.

The next installment will require no knowledge of the telecommunications industry at all.
For those who found this series difficult to follow up to this point, or for those in fields unrelated to telecommunications, you are welcome to start reading from the next installment.

There are no new technical points, so skipping this will not hinder your understanding of Act 2.
Even so, if it helps even one more person feel that 'this methodology also applies to my own workplace,' then I believe it is worth writing.

The first part of Act 2, 'From Pipelines to Multiple Loops,' will follow that.
I hope you will continue to join me.

Notes (reposted)

• This article represents my personal views and does not represent my affiliated organization.

• The products and initiatives mentioned in this series (Windows 365 for Agents, Microsoft Entra Agent ID, Project Opal, Microsoft Scout, etc.) are all based on publicly available information from each company and organization. Windows 365 for Agents was announced as Generally Available (GA) at Build 2026. Some, such as Microsoft Scout (Frontier preview), are in the preview stage, and their names, specifications, and scope of availability are subject to change. Please check official sources for the latest and most accurate information.

• For details on industry standards (3GPP, TM Forum, ITU, ETSI, OWASP, NIST, etc.), the official documents from each organization are the latest and most accurate sources of information.

References (Main Sources)

Microsoft Official Documentation (Microsoft Learn)

• Windows 365 for Agents (Overview) — https://learn.microsoft.com/windows-365/agents/introduction-windows-365-for-agents

• Microsoft Entra Agent ID (What are agent identities) — https://learn.microsoft.com/entra/agent-id/what-are-agent-identities

• Project Opal / Microsoft 365 Copilot Extensibility — https://learn.microsoft.com/microsoft-365-copilot/extensibility/

• Microsoft Scout (Overview) — https://learn.microsoft.com/microsoft-scout/overview

Series Index (Act 1 Complete Edition)

• Part 1: The Wall of Autonomous Operations — Why We Get Stuck at L3

• Part 2: Conditions for Reaching AN Level 4 — The Three-Layer Structure of Observation, Judgment, and Execution

• Part 3: The Void of the Execution Layer — Why the Entire Industry Is Clogged

• Part 4: Head, Hands, and Control — The Three Elements Required for the Execution Layer

• Part 5: The Three-Element Matrix — Re-evaluating Existing Approaches

• Part 6: The New Horizon of the Agentic AI Runtime

• Part 7: Windows 365 for Agent — The 'Foundation for the Execution Layer' the Industry Has Been Waiting For Has Emerged

• Extra Edition (Part 7.5): 95% Is Caused by Human Error — The New Main Battlefield of Agent ID

• Part 8: Project Opal — The Uncharted Territory of the Orchestration Layer

• Part 9: Application to Telco AN ① — Fault Handling Closed-Loop

• Part 10: Application to Telco AN ② — Service Deployment and the Wall of Vendor GUIs

• Part 11: Application to Telco AN ③ — Self-Healing, Self-Optimization, and the Learning Loop

• Part 12: Why Microsoft — The Theory of Structural Advantage

• Part 13: Supplementary Discussion ① — Connecting the TM Forum AN Framework with This Series

• Part 14: Supplementary Discussion ② — The Wall of Control and Its True Height

• Part 15: Supplementary Discussion ③ — The Wall of Observation — From Substrate Observability to Outcome Observability

• Part 16: Supplementary Discussion ④ — The Wall of Defense — Detection and Structural Defense Are Different Jobs

• Part 17 (Act 1 Finale): The Execution Layer of Autonomous Networks, and Beyond (This Article)- Intermission (Part 17.5): You Don't Need Communication Knowledge for Autonomous Operations — I Verified That with 16kg of My Own Body (Next Time)

Act 2 Preview (Cross-Industry Agentic AI Runtime Series)

(Before that, we will have one intermission)

• 3+2 Architecture (Cross-Industry Version)

• Agentic Closed Loop Maturity Model (L0 to L6)

• Detailed Explanation of the 4-Layer Memory (Episodic / Semantic / Procedural / Constitutional)

• Independence of the Governance Loop

• Human Role Shift and Organizational Transformation

• Discussion on AGI Deployability

This article: approx. 7,900 characters / 14-minute read

Author: Hiroyuki Hikino
20 years in the telecommunications industry. 3G/4G/5G mobile core network design / Former Cisco network engineer / Currently at Microsoft (speaking as an individual) / Representative of Diestaday LLC

Please feel free to share your opinions or questions in the article comments or on LinkedIn.

To everyone who has read Act 1 this far, thank you very much.
If you resonated with this series, please 'like' and follow on note, and repost on LinkedIn.
Next time is the intermission, and Act 2 begins after that. I look forward to seeing you there.