[OPS-AREA]Re: [nmrg] Re: [NMOP] Re: [OPSAREA @IETF126] IRTF/IETF OPS Transfer: Status & Exploring Coll aboration Opportunities
"Joe Clarke \(jclarke\)" <[email protected]> Thu, 25 Jun 2026 18:27:26 +0000
| Newsgroups | gmane.ietf.ops,gmane.ietf.opsawg |
|---|---|
| Message-ID | <CH2PR11MB8867DE2EF646093A1B9021ABB8EC2@CH2PR11MB8867.namprd11.prod.outlook.com> |
--===============3044740756372963207==
Content-Language: en-US
Content-Type: multipart/alternative;
boundary="_000_CH2PR11MB8867DE2EF646093A1B9021ABB8EC2CH2PR11MB8867namp_"
--_000_CH2PR11MB8867DE2EF646093A1B9021ABB8EC2CH2PR11MB8867namp_
Content-Type: text/plain; charset="Windows-1252"
Content-Transfer-Encoding: quoted-printable
Chongfeng, I mostly agree with everything you said. That is, I have a simi=
lar understanding. I do have a few comments, though, based on my implement=
ation experience.
Q1. Which AI agents capabilities are suitable for network management?
A: AI agents have the following common capabilities,
- Perception: Processes multi-modal inputs (text, images, audio) and tr=
acks context and state over time.
[JMC] This depends on model. Not all models can handle all types of input.=
And the model capabilities will influence overall cost.
- Planning & Reasoning: Breaks high-level goals into actionable sub-tas=
ks, dynamically re-plans when obstacles arise, and evaluates trade-offs bef=
ore acting.
- Tool Use & Execution: Autonomously calls external tools=97search engi=
nes, APIs, calendars, code interpreters, and databases=97to take real actio=
n in the digital world (e.g., send emails, book flights, query data).
[JMC] More specifically, with network management, tools can access devices =
to get data and adjust configuration. In the former, what the agent can th=
en do with the data it retrieves heavily depends on the model and its train=
ing data. For example, many frontier models generally do well with basic n=
etwork operational data. Adding on to tool use, RAG or purpose-trained or =
tuned models, and you can get even better results from such tools calls.
- Learning & Self-Correction: Detects errors, reflects on outcomes, ite=
rates on solutions, and utilizes both short-term and long-term memory to im=
prove over time.
[JMC] This is where Digital Twin comes into place. You don=92t necessarily=
want a model to =93learn=94 as it=92s going through your production networ=
k. Testing its =93reasoning=94 in a twin environment makes it safer for do=
ing that self-correction and iteration.
All the capabilities above are useful for network management. With the unde=
rground support of LLM models, the following functionalities can be realize=
d by AI agents in the context of network management,
- Intent-based auto-assignment: Automatically routes NOC tickets based o=
n detected intent.
- Intelligent root-cause analysis: Pinpoints fault origins with precisio=
n.
[JMC] Eh, maybe. See above. I=92ve had models go off the rails on hyper-s=
pecific problems where it didn=92t have a lot of relevant training data.
- Actionable troubleshooting guidance: Generates step-by-step plans and=
recommended resolutions.
- AI-driven orchestration: Enables expert-human collaboration and autom=
ated execution of task reassignments/suspensions.
- Continuous network optimization: Delivers ongoing tuning recommendati=
ons for wireless, core, and cloud-network resources across diverse scenario=
s.
- Automated recovery verification: Validates fault resolution by queryi=
ng network and perceptual metrics via LLM-orchestrated API calls.
Q2. What is the impact of an AI network management process to YANG informat=
ion, service, network and data modelling?
A: AI and legcay network management models belong to different domain=
s, the specific manangement modelling is aganostic to the AI network manan=
gent process. In other words, AI system can process any YANG information, s=
ervice, network and data modelling.
[JMC] This has exactly been my experience. I=92ve had good luck with a YAN=
G-centric skill. Where things can get problematic is the token consumption=
XML and JSON produce when using open weight models with tighter context wi=
ndows.
Q3. Is MCP (https://en.wikipedia.org/wiki/Model_Context_Protocol) a suitab=
le protocol between AI agents and IETF network management data sources and =
system components?
A: Yes, as an open standard that connects AI models to external data =
and tools, MCP enables AI assistants to securely read files, query database=
s, and execute actions across different applications. The likelihood of MC=
P being used between between AI agents and IETF network management data sou=
rces and system components is very high.
In addition, other approaches can be used, such as APIs exposed by network =
controller, in this case, AI agents can call the APIs directly for network =
device configuration or data collection.
[JMC] Agree completely. This does work well. And MCP has considerable fra=
mework backing.
Q4, How does an AI agent interact with Network Digital Twin, Intent-Based N=
etworking, SIMAP, Knowledge Graphs and YANG data in Message Brokers?
A: It depends on the specific situation, there should be no one-size-fi=
ts-all solution.
[JMC] See above on a thought on digital twin. And, yeah, I don=92t think w=
e=92ll have a one-size fits all model.
Joe
--_000_CH2PR11MB8867DE2EF646093A1B9021ABB8EC2CH2PR11MB8867namp_
Content-Type: text/html; charset="Windows-1252"
Content-Transfer-Encoding: quoted-printable
<html>
<head>
<meta http-equiv=3D"Content-Type" content=3D"text/html; charset=3DWindows-1=
252">
</head>
<body>
<div style=3D"direction: ltr; font-family: Aptos, Arial, Helvetica, sans-se=
rif; font-size: 12pt; color: rgb(0, 0, 0);">
Chongfeng, I mostly agree with everything you said. That is, I have a=
similar understanding. I do have a few comments, though, based on my=
implementation experience.</div>
<div style=3D"direction: ltr; font-family: Aptos, Arial, Helvetica, sans-se=
rif; font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<div id=3D"mail-editor-reference-message-container">
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
</div>
<div style=3D"padding: 3pt 0in 0in; border-width: 1pt medium medium; border=
-style: solid none none; border-color: rgb(181, 196, 223) currentcolor curr=
entcolor;">
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; text-align: left; font-family: Aptos; font-size: 12pt; color:=
black;">
<br>
</div>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing">Q1. Which A=
I agents capabilities are suitable for network management? </di=
v>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p;A: AI agents have the following common capabilities, </d=
iv>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; - Perception: Processes multi-modal inputs (text, images, audio) and tra=
cks context and state over time.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
[JMC] This depends on model. Not all models can handle all types of i=
nput. And the model capabilities will influence overall cost.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; - Planning & Reasoning: Breaks high-level goals into actionable sub-=
tasks, dynamically re-plans when obstacles arise, and evaluates trade-offs =
before acting.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; - Tool Use & Execution: Autonomously calls external tools=97search e=
ngines, APIs, calendars, code interpreters, and databases=97to take real ac=
tion in the digital world (e.g., send emails, book flights,
query data).</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
[JMC] More specifically, with network management, tools can access devices =
to get data and adjust configuration. In the former, what the agent c=
an then do with the data it retrieves heavily depends on the model and its =
training data. For example, many frontier
models generally do well with basic network operational data. Adding=
on to tool use, RAG or purpose-trained or tuned models, and you can get ev=
en better results from such tools calls.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; - Learning & Self-Correction: Detects errors, reflects on outcomes, =
iterates on solutions, and utilizes both short-term and long-term memory to=
improve over time.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
[JMC] This is where Digital Twin comes into place. You don=92t necess=
arily want a model to =93learn=94 as it=92s going through your production n=
etwork. Testing its =93reasoning=94 in a twin environment makes it sa=
fer for doing that self-correction and iteration.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing">All the cap=
abilities above are useful for network management. With the underground sup=
port of LLM models, the following functionalities can be realized by AI age=
nts in the context of network management,</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; </div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p;- Intent-based auto-assignment: Automatically routes NOC tickets based on=
detected intent.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p;- Intelligent root-cause analysis: Pinpoints fault origins with precision=
.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
[JMC] Eh, <i>maybe</i>. See above. I=92ve had models go off the=
rails on hyper-specific problems where it didn=92t have a lot of relevant =
training data.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p;- Actionable troubleshooting guidance: Generates step-by-step plans=
and recommended resolutions.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p;- AI-driven orchestration: Enables expert-human collaboration and a=
utomated execution of task reassignments/suspensions.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p;- Continuous network optimization: Delivers ongoing tuning recommen=
dations for wireless, core, and cloud-network resources across diverse scen=
arios.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p;- Automated recovery verification: Validates fault resolution by qu=
erying network and perceptual metrics via LLM-orchestrated API calls.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; </div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing">Q2. What is=
the impact of an AI network management process to YANG information, servic=
e, network and data modelling?</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; A: AI and legcay network management models belong to differe=
nt domains, the specific manangement modelling is aganostic to the AI=
network manangent process. In other words, AI system can process
any YANG information, service, network and data modelling. </di=
v>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
[JMC] This has exactly been my experience. I=92ve had good luck with =
a YANG-centric skill. Where things can get problematic is the token c=
onsumption XML and JSON produce when using open weight models with tighter =
context windows.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing">Q3. I=
s MCP (https://en.wikipedia.org/wiki/Model_Context_Protocol) a suitable pro=
tocol between AI agents and IETF network management data sources and system=
components?</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> &nbs=
p; A: Yes, as an open standard that connects AI models to exter=
nal data and tools, MCP enables AI assistants to securely read files, query=
databases, and execute actions across different applications.
The likelihood of MCP being used between between AI agents and IETF networ=
k management data sources and system components is very high. </div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing">In addition=
, other approaches can be used, such as APIs exposed by network controller,=
in this case, AI agents can call the APIs directly for network device conf=
iguration or data collection. </div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
[JMC] Agree completely. This does work well. And MCP has consid=
erable framework backing.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing">Q4, How doe=
s an AI agent interact with Network Digital Twin, Intent-Based Networking, =
SIMAP, Knowledge Graphs and YANG data in Message Brokers?</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing"> A:&n=
bsp; It depends on the specific situation, there should be no one-siz=
e-fits-all solution.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr;">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
[JMC] See above on a thought on digital twin. And, yeah, I don=92t th=
ink we=92ll have a one-size fits all model.</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<div class=3D"ms-outlook-mobile-reference-message skipProofing" style=3D"di=
rection: ltr; font-family: "&quot;Aptos&quot;", Arial, He=
lvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
Joe</div>
</div>
</body>
</html>
--_000_CH2PR11MB8867DE2EF646093A1B9021ABB8EC2CH2PR11MB8867namp_--
--===============3044740756372963207==
Content-Type: text/plain; charset="utf-8"
MIME-Version: 1.0
Content-Transfer-Encoding: base64
Content-Disposition: inline
X19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX18KT1BTLUFSRUEg
bWFpbGluZyBsaXN0IC0tIG9wcy1hcmVhQGlldGYub3JnClRvIHVuc3Vic2NyaWJlIHNlbmQgYW4g
ZW1haWwgdG8gb3BzLWFyZWEtbGVhdmVAaWV0Zi5vcmcK
--===============3044740756372963207==--