| Internet-Draft | AI4AN | July 2026 |
| Clemm & Eckert | Expires 7 January 2027 | [Page] |
This document specifies how the rise of agentic AI and LLMs can impact
and and accelerate the transition towards Intent-Based Networking.
Specifically, it revisits functionality and
liefecycle in IBN, as defined in [RFC9315],
and outlines how agentic AI and LLMs can be leveraged.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 7 January 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Intent-Based Networking (IBN) involves the concept of allowing networks to be managed using "intent", that is, by allowing network operators to specify in a declarative manner sets of operational goals (that a network should meet) and outcomes (that a network is supposed to deliver) without needing to specify how to actually achieve or implement them [RFC9315]. Determining the specific actions that would need to be taken would be up to the IBN itself, whether those actions involve determining configuration parameters and applying them, performing optimizations, monitoring the network and assessing whether additional actions need to be taken, and more. Intent and IBN in many ways complete the vision articulated as part of autonomic networking that involve achieving self-configuration, self-optimization, self-healing, and self-protection while minimizing dependencies on human administrators who are taken out of the control loop. However, even autonomic networks are not clairvoyant but need to be given guidance by administrators. That guidance is characterized as intent [RFC7575].¶
Since Autonomic Networking and IBN were defined, many commercial offerings for intent-based systems have appeared. In many cases, "intent" by those systems has been less abstract in nature but instead involves structured APIs that are used to manage services and connectivity in a network instead of individual network elements on a per-device basis. While useful, there is still a gap between the abstractions provided by those APIs and their underlying system on one hand and the vision of an intent as defined in [RFC9315] on the other hand. For example, operators are still very much responsible to determine strategies to achieve broad outcomes.¶
The recent explosive rise of Large Language Models (LLMs) and agentic AI (roughly understood as systems with LLMs at their core, extended with additional capabilities that allow them to access additional tools, to decompose tasks, to integrate with other applications) holds the promise to potentially change that as it opens up possibilities for the implementation of capabilities in technical systems that until recently did not exist. For example, agentic AI allows for the implementation of chat interfaces that allow users to state their intent in their own words, without requiring them to be proficient in any particular APIs or command interfaces. Importantly, agentic AI may allow to clarify ambiguities, can ask for additional information and details, point out constraints, discuss alternatives, thus providing means for ingesting intent that systems up to this point did not have. Likewise, agentic control loops hold the promise to potentially allow IBNs to intelligently monitor the network to assess outcomes as well as the effectiveness of actions that are taken. Coupled with learning capabilities, IBNs can become even more effective over time and result in continued network optimization and improvement. While some of this promise yet remains to be proven, the potential for it is there and appears to be only a matter of time.¶
The concept of IBN remains very much valid and appears more relevant than ever for the future of networking in light of the new possibilities that can help make it a reality. Given this, it seems reasonable to revisit IBNs, specifically, their functional components as well as the lifecycle reference architecture, and articulate the ways in which agentic AI and LLMs can facilitate and accelerate IBN development and adoption.¶
IBN is subject to several loops, as depicted in the IBN Lifecycle Reference model (Figure 1).¶
User Space : Translation / IBS : Network Ops
: Space : Space
: :
+----------+ : +----------+ +-----------+ : +-----------+
Fulfill |recognize/|---> |translate/|-->| learn/ |-->| configure/|
|generate | | | | plan/ | | provision |
|intent |<--- | refine | | render | : | |
+----^-----+ : +----------+ +-----^-----+ : +-----------+
| : | : |
.............|................................|................|.....
| : +----+---+ : v
| : |validate| : +----------+
| : +----^---+ <----| monitor/ |
Assure +---+---+ : +---------+ +-----+---+ : | observe/ |
|report | <---- |abstract |<---| analyze | <----| |
+-------+ : +---------+ |aggregate| : +----------+
: +---------+ :
The first of these loops occurs between the functions to ingest (recognize, generate) intent, and translate it into specific commands and operations that will be actionable by the underlying network. Agentic AI systems implement similar loops in other areas, for example in program development and coding. Application of Agentic AI to IBN in this area appears thus straightforward and it is expected that corresponding Agentic AI agents cand implement both of these functions.¶
The second loop concerns the "inner" intent control loop between IBS and Network Operations space: learn/plan/render -> configure/provision -> monitor/observe -> analyze/aggregate -> validate -> learn/plan/render (here the loop closes). Agentic AI has inherent translation capabilities that can play an important role in simplifying the implementation of configuration and provisioning systems by determining how to map required actions to device interfaces, ensuring the right commands and APIs are invoked. The same translation capabilities help smoothen over impedance mismatches in the data models exposed by underlying systems. This simplifies many past issues associated with dealing with interface heterogeneity and associated sustaining operations.¶
Agentic AI can also help determine what type of monitoring and measurements will be required to observe the effectiveness of actions taken and assess compliance with intent, and configuration and provisioning functions can be augmented with corresponding capabilities to set up those observations as well. Agentic AI may have less of a role to play in analyzing raw monitoring data as the data velocity in this space is very high. This concerns IPFIX records, IOAM telemetry data, syslog notifications, YANG-Push management data streams. For a network, the volume of these easily reaches into the thousands if not millions of data items per second, outpacing LLM token processing capabilities for the foreseeable future unless other measures are taken to reduce the amount of data by several orders of magnitude. Anomaly detection and trend analysis are areas in which AI-based systems play an important role, but not areas in which Agentic AI based on LLMs are expected to make a considerable impact. However, this can change once raw data is aggregated, filtered, and preprocessed. This preprocessed data can very well serve as input to Agentic AI systems as part of the validation and learning/planning portions of the inner control loop and we expect corresponding solution architectures to emerge.¶
The third loop concerns the "outer" intent control loop that involves abstraction and reporting to keep users informed (and, in a bigger sense, "in the loop" to understand what and how their network is in fact doing). There are many applications of Agentic AI systems to provide decision support, generate smart reports, and highlight relevant information. IBN should not be expected to be any different and Agentic AI is thus a clear candidate to implement the abstraction and assurance functions in the IBN lifecyle.¶
The following sections revisit the corresponding sections in [RFC9315] where IBN functionality is described. For a more detailed discussion of what precisely the functionality entails, please refer to that document. In the following, we look specifically at the ramifications brought about by LLMs and agentic AI for each of those functions.¶
The first set of functions is concerned with having users convey intent to the network respectively to the Intent-Based System (IBS) that communicates on the network's behalf. Using agentic AI underpinned by LLMs can allow for conversational interaction between the user and the IBS. Not only can the user convey intent using plain language, but ambiguities can be resolved, additional context requested, missing information identified, priorities established, tradeoffs presented, all with minimal learning curve on the side of the user. To be effective, the AI agent needs to have access to data about the network, including but not limited to inventory and topology information.¶
Intent translation very closely interacts with intent ingestion; in an agentic AI environment both functions may very well be integrated, not separated. Agentic AI will be able to translate ingested intent into APIs and structured commands, many of which may be device-dependent. In the past, heterogeneity of device interfaces, command interfaces, exposed data models were a major contributing factor to the complexity it took to develop operations support systems. One major value that Agentic AI and LLM technology can provide involves abstracting this heterogeneity and facilitate the application of unified "intent" regardless of the particularities of individual device interfaces, determining what specific APIs to call and what structured commands to apply depending on the specific network device.¶
To the extent that agentic AI can help with planning courses of actions, orchestration of rolling out intent across a network is another function that can benefit from agentic AI, which can offer exceptional planning abilities. One big concern with any orchestration system involves dealing with rainy-day scenarios, for example failures in the ability to apply a particular configuration to a particular device. Orchestration systems need to be able to anticipate, detect, and mitigate such possibilities, including but not limited to rolling the overall network back (or forward) into a defined state, as well as containing any potential negative impact. It will be imperative for any AI agent involved in intent orchestration to have corresponding capabilities, even more critically so as to not require to be exposed to device intrinsics that the IBN is supposed to abstract away from.¶
Monitoring, in the context of intent assurance, involves observing the network, conducting measurements, and collecting telemetry data, in order to provide visibility into the network that is required to ensure that intent is actually complied with and to assess the degree to which desired outcomes are achieved. Some aspects of monitoring are resource-intensive operations that need to be used sparingly. Examples include configuration of measurement probes and generating test traffic (consuming CPU and bandwidth), collecting telemetry data (incurring overhead on production traffic), or continuous streaming of statistics about flows, interfaces, and device state (at a minimum, involving sampling strategies to keep overhead to an acceptable level). Agentic AI can help optimize monitoring strategies by having observations directed at those areas where they matter the most, such as providing answers as to where measurement probes need to be directed at any one point or which aspects are most critical to observe closely and adjust where to probe and what to sample accordingly. Similarly, it can collect additional information or run tests that will be helpful to validate hypotheses about what fine tuning and adjustements maybe needed and facilitate corresponding decisions to do so.¶
Agentic AI is a natural fit to many of the functions that involve abstracting and summarizing data and statistics reported from the network to human users. Coupled with the ability to relate the observations with context and provide analysis to make actionable recommendations, agentic AI can be expected to result in very powerful systems in that area going forward. Previously, the interpretation of data required operator expert knowledge and programming decision support systems to analyze results was a formidable task which will be significantly simplified.¶
TBD¶
draft-cxxx-nmrg-ai4ibn-00: Initial version¶