nmop Q. Sun
Internet-Draft R. Luo
Intended status: Standards Track Y. Fu
Expires: 20 January 2027 C. Xie
China Telecom
19 July 2026
The requirements of Agent Lifecycle Management
draft-sun-nmop-agent-lifecycle-management-00
Abstract
As agents evolve from auxiliary analysis toward autonomous decision-
making and task execution, their operation involves multiple dynamic
factors, including models, knowledges, metrics, tools, permissions,
behaviors, and collaborative relationships. Traditional management
methods are no longer sufficient to ensure their secure and stable
operation. This draft proposes an agent lifecycle management
requirement covering registration and publication, deployment and
operation, continuous operations, capability evolution, deactivation,
and retirement. It analyzes key management requirements related to
assets, capabilities, knowledge, tools, identity and authorization,
operational evaluation, auditing, and version evolution and
retirement, providing the unified management, continuous governance,
and large-scale deployment of agents.
Status of This Memo
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 20 January 2027.
Copyright Notice
Copyright (c) 2026 IETF Trust and the persons identified as the
document authors. All rights reserved.
Sun, et al. Expires 20 January 2027 [Page 1]
Internet-Draft AgentOps July 2026
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.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1. Scope . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2. Requirements Language . . . . . . . . . . . . . . . . . . 6
1.3. Definition and Terminology . . . . . . . . . . . . . . . 6
2. Why the Agent Lifecycle Management is important? . . . . . . 6
2.1. Network Operation Agents Are Evolving from Decision Support
to Autonomous Execution, Increasing Operational Risks . . 6
2.2. Network Operation Agents Continuously Evolve, Making
One-Time Management Approaches Insufficient . . . . . . . 7
2.3. Network Operation Agents Depend on Multiple Capability
factors and Require Unified Management . . . . . . . . . 7
2.4. Large-Scale Deployment of Network Operation Agents Requires
a Unified Operational Governance Mechanism . . . . . . . 8
2.5. Network Operation Agents Should Continuously Create the
Value of Operation . . . . . . . . . . . . . . . . . . . 8
3. Requirements for Agent Lifecycle Management . . . . . . . . . 8
3.1. Agent Asset and Accountability Management . . . . . . . . 9
3.2. Agent Capability Composition and Configuration
Management . . . . . . . . . . . . . . . . . . . . . . . 9
3.3. Network Performance Parameters, Knowledge, Ontology, and
Memory Management . . . . . . . . . . . . . . . . . . . . 10
3.4. Tool, Skill, and Collaboration Relationship Management . 11
3.5. Identity, Authorization, and Trust Relationship
Management . . . . . . . . . . . . . . . . . . . . . . . 12
3.6. Quality, Effectiveness, and Performance Management . . . 13
3.7. Runtime Behavior, Reliability, and Cost Management . . . 13
3.8. Security, Risk, Compliance, and Audit Management . . . . 14
3.9. Version Change, Evolution, and Retirement Management . . 15
4. Security Considerations . . . . . . . . . . . . . . . . . . . 16
4.1. Identity and Trust Security . . . . . . . . . . . . . . . 16
4.2. Permission and Authorization Security . . . . . . . . . . 16
4.3. Data and Knowledge Security . . . . . . . . . . . . . . . 16
4.4. Tool and Supply Chain Security . . . . . . . . . . . . . 17
4.5. Behavior and Execution Security . . . . . . . . . . . . . 17
4.6. Lifecycle Security . . . . . . . . . . . . . . . . . . . 17
4.7. Observability and Audit Security . . . . . . . . . . . . 17
5. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 17
Sun, et al. Expires 20 January 2027 [Page 2]
Internet-Draft AgentOps July 2026
6. Normative References . . . . . . . . . . . . . . . . . . . . 17
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . 18
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 18
1. Introduction
As large language model technologies continue to evolve from content
generation toward task planning, tool invocation, and autonomous
execution, agents operating within network environments—hereinafter
referred to as agents—are gradually becoming important execution
entities in network operations and maintenance, service operations,
and network management. Agents are capable of invoking network
management systems, operations and maintenance tools, digital twin
platforms, knowledge bases, and model services to autonomously
perform network operation tasks such as fault analysis, alarm
correlation analysis, configuration generation, inspection analysis,
and capacity optimization. They are also increasingly assuming some
of the decision-making and execution responsibilities traditionally
undertaken by human operators. Unlike conventional automation
programs that operate according to predefined logic, agents can
autonomously interpret tasks, plan execution procedures, select
appropriate tools, invoke relevant resources, and complete task
execution in accordance with operational objectives. Some agents are
further equipped with persistent memory, experience accumulation, and
multi-agent collaboration capabilities.
As the autonomy and environmental adaptability of network operation
agents continue to increase, they can no longer be regarded merely as
static software programs. Rather, they have evolved into composite
digital entities consisting of multiple interdependent elements,
including models, prompts, network knowledge, runtime memory, tool
plug-ins, Model Context Protocol (MCP) services, workflows,
authorization policies, and execution environments. During the
execution of network operations and maintenance tasks, an agent may
dynamically adjust its execution strategy according to the current
network state, service context, device capabilities, and the status
of external tools. Consequently, even when processing the same
operational task, different reasoning processes and execution
outcomes may arise as a result of variations in model versions,
knowledge content, runtime environments, or collaborating entities.
At the same time, network operation agents may act on behalf of
operations personnel or network management systems to access network
resources, invoke operations and maintenance tools, execute
configuration changes, and collaborate with other agents in
completing complex network operation tasks. As the number and scale
of network operation agents continue to grow, the models, knowledge,
tools, network metric, permissions, and collaborative relationships
Sun, et al. Expires 20 January 2027 [Page 3]
Internet-Draft AgentOps July 2026
involved in their operation will also change continuously. In the
absence of a unified management mechanism, a range of issues may
arise, including limited visibility of agent assets, unclear
accountability, poorly defined capability boundaries, excessive
retention of permissions, insufficient observability of operational
behavior, uncontrolled version evolution, and failure to reclaim
resources following agent deactivation. These issues may, in turn,
compromise network security and the stability of the overall network
operations and maintenance.
It is therefore necessary to pay more attention to the lifecycle
management problems for agents operating in network environments,
covering the entire process from registration and publication,
deployment and operation, and continuous management to capability
evolution, deactivation, and retirement. It should provide unified
management of the key objects that continuously change throughout
agent operation, ensuring that network operation agents remain
discoverable, configurable, observable, assessable, controllable, and
traceable, while continuously, securely, and reliably supporting the
execution of network operations and maintenance tasks.
1.1. Scope
The Agent focused in this draft is for network operation and
management — how AI agents are configured, governed, and monitored,
not AI applied to networking more broadly. Agent Lifecycle
Management defined in this draft refers to a set of management
policies, operational processes, and technical capabilities
established for agents operating within network environments. It
covers the entire lifecycle of an agent, from registration and
publication, deployment and operation, and continuous operations to
capability evolution, deactivation, and retirement. Focusing on core
management objects—including agent identity, assets, capability
configurations, knowledge, memory, tools, workflows, metrics,
permissions, runtime behavior, versions, and collaborative
relationships—it establishes mechanisms for unified governance,
continuous observation, operational evaluation, risk control, change
management, and decommissioning and resource reclamation. Its
objective is to ensure that agents remain discoverable, configurable,
observable, assessable, controllable, and traceable throughout their
operational lifecycles, while enabling them to perform network-
related tasks securely, reliably, and efficiently.
Agent lifecycle management may considered by the following three
components:
Sun, et al. Expires 20 January 2027 [Page 4]
Internet-Draft AgentOps July 2026
• End-to-End Operational Management: Agent lifecycle management
encompasses the complete operational process of an agent, including
registration and publication, deployment and operation, continuous
operations, capability evolution, deactivation, and retirement. It
is not limited to runtime monitoring or fault handling after an agent
has been deployed. Lifecycle management requires the continuous
maintenance and tracking of agent state transitions across different
operational stages, ensuring that agents can operate reliably as
network environments, service requirements, and operations and
maintenance tasks evolve. Once an agent is withdrawn from operation,
its identity must be deregistered, its permissions revoked, and its
associated resources released, thereby establishing a complete
closed-loop operational management process.
• Multi-factor Collaborative Management: The operational capabilities
of a network operations agent are jointly determined by multiple
constituent factors. Accordingly, lifecycle management must address
not only the agent software itself, but also its models, prompts,
knowledge bases, memory, tools, MCP services, workflows, network
metrics, identities, permissions, and runtime environments. Changes
to any of these factors may affect the agent’s ability to understand
network conditions, the manner in which it performs tasks, and the
risks associated with its operation. Agent lifecycle management must
therefore establish a unified management mechanism for all
constituent factors and maintain their consistency, integrity, and
traceability.
• Continuous Operations and Dynamic Governance: The publication of an
agent does not mark the completion of lifecycle management; rather,
it marks the beginning of the continuous operational phase. As
network topologies change, device versions are upgraded, knowledge
continues to accumulate, models are iteratively updated, and
operations and maintenance strategies are adjusted, the capabilities
of an agent will continue to evolve. During operation, agents may
also experience behavioral drift, tool failures, outdated knowledge,
resource anomalies, and security risks. Lifecycle management must
therefore continuously conduct runtime observation, performance
evaluation, capability calibration, risk governance, and version
evolution. Configurations, permissions, knowledge, and control
policies should be dynamically adjusted according to the agent’s
operational state, ensuring that the agent remains secure, reliable,
and controllable throughout its entire operational lifecycle.
Sun, et al. Expires 20 January 2027 [Page 5]
Internet-Draft AgentOps July 2026
1.2. Requirements Language
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
capitals, as shown here.
1.3. Definition and Terminology
• AI：Artificial intelligence
• LLM：Large Language Model
• MCP：Model Context Protocol
• NMS：Network Management System
• OSS：Operations Support System
2. Why the Agent Lifecycle Management is important?
As agents operating within network environments increasingly
undertake tasks such as network monitoring, fault diagnosis,
configuration changes, inspection analysis, capacity optimization,
and network scheduling, they are evolving from auxiliary analytical
tools into important execution entities in network operations.
Network operation agents are capable of autonomously planning tasks,
invoking models, accessing network resources, performing operations
and maintenance actions, and collaborating with other agents to
complete complex tasks. Their operational behavior may therefore
directly affect network performance and service continuity.
Conventional software operations and model management approaches are
primarily designed for static applications and are insufficient to
address the continuous evolution, autonomous decision-making, and
dynamic collaboration of network operation agents. It is therefore
necessary to establish a management mechanism that spans the entire
operational lifecycle of network operation agents.
2.1. Network Operation Agents Are Evolving from Decision Support to
Autonomous Execution, Increasing Operational Risks
Traditional network intelligence systems primarily provide functions
such as alarm analysis, fault prediction, and optimization
recommendations, while final decisions and execution remain the
responsibility of operations personnel. As agent capabilities
continue to advance, network operation agents are increasingly able
to autonomously perform tasks such as fault localization,
Sun, et al. Expires 20 January 2027 [Page 6]
Internet-Draft AgentOps July 2026
configuration generation, tool invocation, policy execution, and
cross-system coordination. They are therefore evolving from
auxiliary analysis toward autonomous execution. The enhancement of
agent execution capabilities means that their influence is no longer
limited to information generation but extends directly to network
operation processes. Errors in task planning, abnormal tool
invocation, or inappropriate execution strategies may result in
incorrect configuration deployment, network service interruptions, or
resource scheduling anomalies. Agent lifecycle management is
therefore required to continuously safeguard the operational security
and execution reliability of network agents.
2.2. Network Operation Agents Continuously Evolve, Making One-Time
Management Approaches Insufficient
The operational capabilities of network agents are jointly influenced
by multiple factors, including models, knowledge, prompts, tools,
network environments, and service policies. Their behavior may
continuously change as models are upgraded, knowledge is updated,
network topologies evolve, and tool capabilities are adjusted.
Accordingly, the operational state of a network operation agent does
not remain fixed after deployment. Reliance solely on pre-deployment
testing or version acceptance cannot ensure long-term operational
stability. Lifecycle management must extend throughout the agent’s
operational process and continuously support runtime observation,
capability evaluation, and dynamic optimization, ensuring that the
agent remains aligned with the evolving network operating
environment.
2.3. Network Operation Agents Depend on Multiple Capability factors and
Require Unified Management
The execution of network operations and maintenance tasks by network
agents typically depends on the coordinated use of foundation models,
network knowledge bases, digital twin platforms, monitoring systems,
task ticket systems, network controllers, MCP services, and other
agents. Changes to any of these constituent elements may affect the
agent’s analytical results and execution behavior. As the scale of
network operation agent deployment increases, complex dependencies
emerge among models, knowledge, tools, network metrics, and
collaborative relationships. In the absence of unified lifecycle
management, issues such as unclear capability boundaries,
configuration drift, tool failures, and collaboration anomalies may
arise. Network agents and their associated capability elements must
therefore be treated as integrated management objects and maintained
continuously through a unified management mechanism.
Sun, et al. Expires 20 January 2027 [Page 7]
Internet-Draft AgentOps July 2026
2.4. Large-Scale Deployment of Network Operation Agents Requires a
Unified Operational Governance Mechanism
As the number of network operations agents continues to grow,
different technical domains, platforms, and vendors may independently
develop and deploy various types of agents that jointly participate
in network operations. Without unified management, this may result
in duplicated agent development, unclear accountability, uncontrolled
identities and permissions, limited visibility into operational
states, and prolonged occupation of network and computing resources.
A unified governance mechanisms therefore required to cover the
entire lifecycle of agent assets, identities, permissions,
operational states, version evolution, decommissioning, and resource
reclamation. It should enable unified onboarding, observation, and
operation of network agents, thereby supporting the coordinated
operation of agents at scale.
2.5. Network Operation Agents Should Continuously Create the Value of
Operation
The objective of deploying network operation agents is not merely to
complete their development, but more importantly to continuously
improve network operations efficiency, reduce fault recovery time,
enhance network performance, and decrease reliance on manual
intervention. Lifecycle management should therefore extend beyond
the deployment stage and continuously evaluate agent performance,
resource consumption, operational costs, and business contributions.
Through continuous evaluation and dynamic optimization throughout the
operational lifecycle, issues such as capability degradation,
outdated knowledge, excessive operational costs, and declining
business value can be identified in a timely manner. The operational
capabilities of network operation agents can then be continuously
optimized to ensure that they remain secure, stable, and efficient
while consistently contributing to the improvement of network
operations and maintenance capabilities.
3. Requirements for Agent Lifecycle Management
As agents increasingly undertake operations and maintenance tasks
such as network monitoring, fault diagnosis, configuration changes,
inspection analysis, capacity optimization, and network scheduling,
the scope of lifecycle management is no longer limited to the agents
themselves. It has expanded to encompass multiple dimensions,
including identity, capabilities, knowledge, tools, behavior,
resources, and security. Throughout different lifecycle stages,
agents continuously undergo capability evolution, knowledge updates,
tool expansion, and changes in operational status. Agent lifecycle
management must therefore provide unified management of the key
Sun, et al. Expires 20 January 2027 [Page 8]
Internet-Draft AgentOps July 2026
objects that continuously change during agent operation, ensuring
that agents remain identifiable, configurable, observable,
controllable, and traceable throughout their entire lifecycles.
3.1. Agent Asset and Accountability Management
The primary objective of agent asset and accountability management is
to determine which agents are operating within the network, what
tasks they are responsible for, who is accountable for them, and how
they are associated with network resources and systems. The relevant
management information typically includes the agent’s unique
identifier, name, source, functional role, technical domain,
deployment location, operational status, accountable entity, and
relationships with network devices, service systems, and other
agents. In network operations and maintenance scenarios, an agent
possesses not only the general attributes of a software asset but may
also be characterized through five key elements: role
responsibilities, runtime environment, behavioral constraints, domain
knowledge, and tool-related skills. Collectively, these elements
define the agent’s role, capability boundaries, and execution methods
within the network operations and maintenance system, thereby
facilitating a consistent understanding of the functions and
responsibilities of different agents. As an increasing number of
agents are developed by different departments, platforms, and
vendors, major challenges include inconsistent agent identification,
limited exchangeability of asset information, unclear accountability
relationships, and difficulties in cross-platform discovery and
onboarding.
* Gap analysis for IETF standards:
1.It may need to establish requirements for unified agent asset
onboarding and management, covering unique identifiers, names,
sources, functional roles, technical domains, deployment
locations, operational states, accountable entities, and
relationships with network devices, service systems, and other
agents. 2.Define the five key elements of an agent to describe its
role, capability boundaries, and execution methods within the
network operations and maintenance system.
3.2. Agent Capability Composition and Configuration Management
The capabilities of a network operation agent are not derived from a
single model. Rather, they are jointly determined by multiple
elements, including LLM models, prompts, knowledge bases, memory,
tools, workflows, runtime policies, and model parameters. When an
agent performs tasks such as network fault diagnosis, configuration
changes, and inspection analysis, its capability boundaries and
execution results depend on the coordinated operation of these
Sun, et al. Expires 20 January 2027 [Page 9]
Internet-Draft AgentOps July 2026
constituent elements. Agent lifecycle management must therefore
manage not only the agent itself but also the composition and
configuration state of its capabilities on a continuous basis.
Network operations environments are characterized by continuous
device upgrades, rapid knowledge updates, ongoing expansion of tool
capabilities, and dynamic adjustments to operational policies.
Consequently, the constituent elements of an agent also change
continuously. Managing only software code or model versions, without
unified management of configuration elements such as prompts,
knowledge, tools, workflows, and runtime policies, makes it difficult
to accurately reproduce an agent’s capability state or analyze the
effects of capability changes on network operations tasks. A unified
management mechanism covering agent capability composition and
configuration elements is therefore required to ensure that
capability composition can be formally described, configuration
states can be tracked, and capability changes can be audited.
* Gap analysis for IETF standards: It may need to establish unified
definitions and management mechanisms for the elements that
constitute and configure agent capabilities.
3.3. Network Performance Parameters, Knowledge, Ontology, and Memory
Management
Network performance metrics, knowledge, ontologies, and memory
collectively constitute the information foundation upon which network
agents perform tasks. Network performance parameters represent the
operational state of the network; ontologies define network entities
and their semantic relationships; knowledge provides operational
rules and expert experience; and memory retains contextual
information and historical experience generated during agent
operation. Together, these elements support agents in performing
complex tasks such as network fault diagnosis, root-cause analysis,
and configuration optimization. As network environments and services
continue to evolve, data, knowledge, ontologies, and memory also
change continuously and must therefore be maintained as important
lifecycle management objects.
Network operations and maintenance scenarios are characterized by
heterogeneous data sources, frequent knowledge updates, and complex
relationships among network entities. In the absence of unified
management, problems may arise, including inconsistent data
provenance, incompatible ontology definitions, outdated knowledge,
and memory contamination. Such problems may impair an agent’s
ability to understand network conditions and effectively perform
tasks. It is therefore necessary to establish a unified management
mechanism for data, knowledge, ontologies, and memory, enabling data
provenance to be traced, knowledge to be updated, ontologies to be
Sun, et al. Expires 20 January 2027 [Page 10]
Internet-Draft AgentOps July 2026
shared, and memory to be governed. This mechanism should provide a
consistent and trustworthy knowledge foundation for the continuous
operation of network agents.
* Gap analysis for IETF standards: 1.It may need to establish a
unified management mechanism for Network performance metrics,
knowledge, ontologies and related data models and memory,
including unified definitions of Network performance metrics,
knowledge, and ontologies, as well as relevant ontology models.
3.4. Tool, Skill, and Collaboration Relationship Management
The ability of network operation agents to evolve from information
analysis toward autonomous task execution depends primarily on their
ability to invoke network tools, apply specialized skills, and
collaborate with other agents in completing complex operations and
maintenance tasks. Tools, skills, and collaborative relationships
therefore define the capability boundaries of an agent and constitute
important management objects within agent lifecycle management.
Lifecycle management must continuously maintain the tools, skills,
and collaborative relationships available to an agent. These may
include network management systems, operations support systems,
monitoring platforms, configuration centers, digital twin platforms,
task tickets systems, network controllers, and various tool
interfaces based on the Model Context Protocol. It must also
maintain specialized skills and invocation relationships associated
with fault diagnosis, alarm analysis, configuration changes,
inspection, and optimization.
As network operations increasingly evolve toward multi-agent
collaboration, a complex operations and maintenance task may require
the coordinated participation of multiple agents. Such agents may
form complex relationships involving task decomposition, capability
invocation, information sharing, and collaborative execution.
Without unified management, issues may arise, including
inconsistencies in tool capabilities, interface invocation failures,
invalid collaborative relationships, and unclear capability
boundaries. A unified management mechanism for tools, skills, and
collaborative relationships is therefore required to continuously
maintain tool versions, interface status, capability dependencies,
and collaboration relationships, thereby ensuring that network agents
can perform cross-system collaborative tasks reliably and
efficiently.
Sun, et al. Expires 20 January 2027 [Page 11]
Internet-Draft AgentOps July 2026
* Gap analysis for IETF standards: 1.It may need to establish a
unified management mechanism for tools, skills, and collaboration
relationships, including standardized definitions of interface
status, capability dependencies, and inter-agent collaboration
relationships.
3.5. Identity, Authorization, and Trust Relationship Management
As network agents increasingly participate in network configuration,
resource management, and automated operations, they can no longer be
regarded merely as conventional software programs. Rather, they
function as digital subjects capable of performing network tasks on
behalf of operations personnel, service systems, or other agents.
Agent lifecycle management must therefore establish a unified
identity, authorization, and trust management mechanism to ensure
that agents possess explicit identities, legitimate permissions, and
trustworthy execution capabilities throughout their operation. Such
management must continuously maintain agent identity information,
identity credentials, accountable entities, authorization sources,
access permissions, operational scope, and trust relationships,
ensuring that every resource access request, tool invocation, and
operations action can be attributed to a specific agent and its
authorization source.
The responsibilities and risk levels of network operations agents
vary considerably, and their access privileges and execution scopes
should therefore differ accordingly. For example, an inspection
agent may be permitted only to query network status, whereas a
configuration change agent may be authorized to modify network
configurations. In the absence of unified identity, authorization,
and trust management, problems such as identity sharing, excessive
authorization, unclear accountability, and unauthorized agent
execution may arise. It is therefore necessary to establish a
management mechanism covering the entire process of agent identity
creation, authorization, modification, suspension, and revocation.
Agent permissions and trust relationships should be dynamically
adjusted according to lifecycle status to ensure the continuous,
secure, and trustworthy operation of network agents.
* Gap analysis for IETF standards: 1.It may need some operational
security best current practices.
Sun, et al. Expires 20 January 2027 [Page 12]
Internet-Draft AgentOps July 2026
3.6. Quality, Effectiveness, and Performance Management
As network agents increasingly undertake production tasks such as
fault diagnosis, configuration changes, network optimization, and
automated remediation, their operational quality directly affects
network stability and operations efficiency. Agent lifecycle
management must therefore assess not only whether an agent can
complete a task, but also the quality of task execution, operational
effectiveness, and business value delivered by the agent. Lifecycle
management should continuously maintain quality indicators for
different stages of agent operation, including task completion rate,
fault localization accuracy, tool invocation success rate, automated
remediation success rate, response latency, stability, and
robustness. These indicators should be combined with measures of
network performance, operational efficiency, and resource consumption
to continuously evaluate agent performance.
Because network environments, model capabilities, knowledge content,
and tools change continuously, agent performance may also vary over
time. Pre-deployment testing alone is therefore insufficient to
guarantee long-term reliability. A continuous evaluation mechanism
spanning the entire operational lifecycle of an agent is required to
support dynamic monitoring of operational quality, continuous
validation of capabilities, and ongoing assessment of business value.
This mechanism should ensure that agents remain stable, efficient,
and trustworthy while continuously contributing to the improvement of
network operations and maintenance capabilities.
* Gap analysis for IETF standards: 1.It may need to define quality
metrics to different stages of agent operation, and establish an
agent performance evaluation mechanism incorporating indicators
such as network performance, operations and maintenance
efficiency, and resource consumption. （This work related to
evaluate the performance of AI agents ). The IETF BMWG has active
work on AI benchmarking that is directly relevant; work in this
area should coordinate with and reference that effort.
3.7. Runtime Behavior, Reliability, and Cost Management
NNetwork agents possess autonomous planning, reasoning, decision-
making, and execution capabilities. Their operation involves not
only task inputs and result outputs, but also task interpretation,
reasoning processes, tool invocation, multi-agent collaboration, and
execution feedback. Agent lifecycle management must therefore
continuously maintain information relating to runtime behavior,
service reliability, and resource consumption to ensure that agents
can perform network operations and maintenance tasks reliably,
securely, and efficiently. Lifecycle management should continuously
Sun, et al. Expires 20 January 2027 [Page 13]
Internet-Draft AgentOps July 2026
collect information concerning task requests, contextual data,
reasoning processes, tool invocations, execution results, anomalous
events, and operational states. This information should be combined
with indicators such as response latency, task success rate, token
consumption, number of model invocations, and resource utilization to
create a comprehensive behavioral profile covering the entire agent
execution process.
Network operations and maintenance tasks are characterized by
stringent real-time requirements, continuous operation, and cross-
system collaboration. An anomaly occurring at any stage may affect
the entire operational workflow. It is therefore necessary to
establish an observability mechanism covering the complete agent
execution process, ensuring that agent behavior can be recorded,
execution processes can be traced, causes of anomalies can be
identified, resource consumption can be measured, and operational
costs can be quantified. Agent reliability should also be
continuously evaluated to support operational optimization, resource
scheduling, and cost control.
* Gap analysis for IETF standards: It may need to define an end-to-
end observability mechanism covering the complete execution
process of agents, including the observability of Model Context
Protocol (MCP) services, LLM models, and agents.
3.8. Security, Risk, Compliance, and Audit Management
As network operation agents increasingly participate in critical
operations and maintenance tasks such as network configuration,
resource scheduling, and fault recovery, their security risks have
expanded beyond conventional software security to encompass models,
data, tools, identities, and multi-agent collaboration. Agent
lifecycle management must continuously maintain information
concerning agent security boundaries, risk levels, permission scopes,
tool authorizations, sensitive data access, security policies, and
audit records. Key events occurring throughout agent operation must
be continuously recorded to ensure that agents perform network tasks
securely and reliably within their authorized scope.
During operation, network agents may be exposed to emerging risks
such as prompt injection, knowledge poisoning, tool misuse,
unauthorized access, incorrect configuration deployment, and risk
propagation across multiple agents. A security governance mechanism
spanning the entire agent lifecycle is therefore required to support
continuous risk identification, dynamic adjustment of security
policies, anomalous behavior detection, and end-to-end audit tracing.
When an agent performs a high-risk network operation, the basis for
the action, authorization source, scope of impact, and accountable
Sun, et al. Expires 20 January 2027 [Page 14]
Internet-Draft AgentOps July 2026
entity should be explicitly identified. The complete execution
process should support traceability and accountability auditing to
ensure the secure and compliant operation of network agents.
* Gap analysis for IETF standards: It may need some operational
security best current practices.
3.9. Version Change, Evolution, and Retirement Management
Network operation agents are continuously evolving entities whose
capabilities change as foundation models, knowledge bases, tools,
prompts, and operations policies are updated. Agent lifecycle
management must therefore address not only the operational status of
an agent but also its version evolution, capability changes, and
retirement process, ensuring that every change is identifiable,
verifiable, and traceable. Lifecycle management should maintain
agent version information, change descriptions, scopes of impact,
evaluation results, release status, and relationships among versions.
It should also continuously record changes to key capability
elements, including models, knowledge, tools, and configurations.
As network devices are upgraded, service requirements change, and
operations capabilities continue to evolve, some agents may be
withdrawn from production environments because their capabilities
have been replaced, their corresponding services have been adjusted,
or they have remained inactive for extended periods. Agent
retirement involves more than terminating operation. It also
requires identity deregistration, permission revocation, tool
disassociation, knowledge and memory cleansing, termination of
collaboration relationships, and archival of operational data. A
lifecycle management mechanism covering version release, upgrade,
modification, replacement, deactivation, and retirement is therefore
required to support continuous agent evolution and secure withdrawal,
while maintaining consistency between operational status and
capability configuration.
* Gap Analysis: 1.It may need to establish a lifecycle management
mechanism covering the complete processes of version release,
upgrade, modification, replacement, deactivation, and retirement.
Section 7 of [RFC 9222] addresses upgrade and lifecycle management
for autonomous agents and is directly relevant existing work in
this area,
Sun, et al. Expires 20 January 2027 [Page 15]
Internet-Draft AgentOps July 2026
4. Security Considerations
Agent lifecycle management should address not only the continuous
evolution of agent capabilities, but also the security of agents
throughout their entire operational lifecycles. Network agents
possess capabilities for autonomous planning, reasoning, tool
invocation, and task execution. They may also act on behalf of users
or systems to access network resources, invoke service interfaces,
and perform operations and maintenance actions. Consequently, their
security boundaries have expanded beyond conventional software
security to encompass identity, authorization, data, knowledge,
tools, collaboration, and runtime behavior. Agent lifecycle
management should therefore establish continuous governance
mechanisms around critical security objects, ensuring that agents
remain trustworthy, controllable, and auditable throughout their
operation.
4.1. Identity and Trust Security
As independent digital subjects, network agents must possess unique
identities and trustworthy identity authentication capabilities.
Lifecycle management should address security issues such as identity
forgery, identity impersonation, cross-domain authentication, and the
establishment of trust relationships among agents. It should ensure
that agent identities remain authentic and trustworthy throughout the
entire lifecycle and that stable trust relationships can be
established and maintained.
4.2. Permission and Authorization Security
Network operatoin agents may access network resources and perform
operations and maintenance actions; therefore, their permissions must
be commensurate with the tasks they undertake. Lifecycle management
should continuously address issues such as excessive authorization,
permission misuse, invalid delegation relationships, and the
prolonged retention of highly privileged agents. Agent authorization
should consistently comply with the principle of least privilege.
4.3. Data and Knowledge Security
During operation, agents continuously access network data, knowledge
bases, and long-term memory. The security of these information
assets involves not only data confidentiality, but also knowledge
integrity, data authenticity, and memory trustworthiness. Lifecycle
management should address risks such as sensitive data leakage,
knowledge poisoning, training data tampering, residual long-term
memory, and untrustworthy data provenance.
Sun, et al. Expires 20 January 2027 [Page 16]
Internet-Draft AgentOps July 2026
4.4. Tool and Supply Chain Security
Network operation agents depend on a wide range of external tools,
Model Context Protocol services, models, plug-ins, and third-party
capabilities. Their security risks therefore extend across the
entire agent supply chain. Lifecycle management should continuously
assess tool trustworthiness, interface security, model provenance,
plug-in integrity, and operational risks arising from changes to
external capabilities.
4.5. Behavior and Execution Security
Because agents possess autonomous reasoning and execution
capabilities, their behavior may drift as a result of changes in
context, prompts, models, or knowledge. Lifecycle management should
continuously address operational security risks such as prompt
injection, malicious instructions, inappropriate tool invocation,
incorrect configuration deployment, and the propagation of erroneous
behavior across multiple agents. Appropriate controls should be
established to ensure that agent behavior remains aligned with
intended objectives and operational constraints.
4.6. Lifecycle Security
New security risks may arise at every stage of a network agent’s
lifecycle, including registration, operation, upgrade, modification,
deactivation, and retirement. Lifecycle management must ensure that
model upgrades, knowledge updates, tool changes, permission
adjustments, and agent retirement processes are verifiable,
auditable, and traceable. It is necessary to prevent residual
identities, unreclaimed permissions, retained data, and uncontrolled
capabilities resulting from inadequate lifecycle management.
4.7. Observability and Audit Security
Critical events throughout the agent lifecycle should be continuously
recorded and traced, including identity changes, permission
adjustments, tool invocations, model updates, knowledge updates,
runtime anomalies, and collaborative interactions. Agent lifecycle
management should establish a unified security audit mechanism to
support operational analysis,accountability tracing, incident
investigation, and security forensics.
5. IANA Considerations
This document has no IANA actions.
6. Normative References
Sun, et al. Expires 20 January 2027 [Page 17]
Internet-Draft AgentOps July 2026
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/info/rfc2119>.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, <https://www.rfc-editor.org/info/rfc8174>.
[RFC9222] Carpenter, B., Ciavaglia, L., Jiang, S., and P. Peloso,
"Guidelines for Autonomic Service Agents", RFC 9222,
DOI 10.17487/RFC9222, March 2022,
<https://www.rfc-editor.org/info/rfc9222>.
Acknowledgements
TBD
Authors' Addresses
Qiong Sun
China Telecom
Beijing
China
Email: sunqiong@chinatelecom.cn
Runze Luo
China Telecom
Beijing
China
Email: luorz4@chinatelecom.cn
Yu Fu
China Telecom
Beijing
China
Email: fuy44@chinatelecom.cn
Chongfeng Xie
China Telecom
Beijing
China
Email: xiechf@chinatelecom.cn
Sun, et al. Expires 20 January 2027 [Page 18]