Video summary

Agentic AI in Telco (Webinar)

Main summary

Key takeaways

Technology

Webinar focus: Agentic AI in Telco

The speaker frames agentic AI as the next step beyond GenAI/chatbots toward autonomous, goal-driven operations in the telecom domain. This is particularly aligned with:

  • TM Forum autonomy levels (L0–L5)
  • Intent-driven operations
  • “Lights-out” / Dark Knock network operations

Speaker background / source

  • Karim Rabia — Principal Architect at Red Hat
  • Telecom experience across:
    • Mobile core
    • OSS / cloud orchestration
    • Participation in standards/working groups
  • Currently pursuing a master’s in AI

Key technological concepts, definitions, and differentiators

1) GenAI vs. agentic AI (core distinction)

  • GenAI/chatbot
    • Typically provides a “journey” (instructions/checklists) in response to a request.
  • Agentic AI
    • Breaks a goal into multi-step tasks
    • Executes those tasks
    • Adapts when dependencies fail
      • Example: visa denied → reschedule flight

Main point: agentic AI aims to fulfill the target/goal, not just generate steps.

2) What “agentic AI / AI agent” means (as described)

The term is presented as a non-standard buzzword, but the webinar consolidates common attributes from vendors/academia:

  • Goal-oriented (goal, not just a task)
  • Coordination via orchestration

  • Loop-based behavior: perceive → reason → plan → act → reflect (or observe)

  • Uses data + tools Agents take actions, so they must integrate with tools/APIs.

  • Maintains state/memory across multi-step interactions

Academia criteria highlighted:

  • Sets a goal
  • Selects/invokes tools to pursue the goal
  • Maintains memory/history/state
  • Completes work fully autonomously

Agent architecture: the “Perceive–Reason–Act–Observe” loop (with telco example)

Perceive (build context)

Telco inputs referenced as context sources:

  • Alarms
  • KPIs
  • Topology
  • Inventory
  • Monitoring feeds via APIs
  • Runbooks/documents via RAG
  • Session and long-term memory (history of prior incidents/procedures)

Protocols/tools mentioned:

  • MCP (Model Context Protocol) to connect to external tools/data/inventories/repositories/etc.

Reason (LLM as the reasoning engine)

  • The LLM / foundation model performs reasoning using the constructed context.
  • The agent may request additional signals (e.g., “pull another KPI”).
  • Reasoning correlates signals and supports:
    • Diagnosis / RCA (root cause analysis)

Act (execute actions through tools)

  • Actions are implemented via:
    • tool/function calls
    • automation/provisioning controllers
  • Skills (procedural knowledge) provide domain-specific workflow steps
    • Example: fault triage/rollback/provisioning playbooks
  • Gated actions may require human-in-the-loop (HITL) approval if high impact.
  • Example action: rollback a bad configuration parameter causing KPI degradation.

Observe (verify results and loop)

  • After acting, the agent checks whether KPIs return to target baseline.
  • It writes results back into memory and either:
    • stops if the goal is met, or
    • loops if not.

Telco momentum and standards alignment (market + autonomy ladder)

1) Adoption claims and autonomy level trend

  • Cites NVIDIA “State of AI” (2024–2026 releases):
    • 97% of telecom professionals using AI, mainly in operations
  • Notes increasing operator statements claiming L4 autonomy adoption (nearly autonomous operations).
  • Mentions telco emphasis around:
    • TM Forum (autonomy framework and projects)
    • 3GPP / 6G (intent-driven management/orchestration and “agentic AI in 6G use cases”)

2) Evolution to autonomy (timeline of enablers)

  • 2017: Transformer paper (“Attention Is All You Need”) → foundation for LLMs
  • 2017 (Telco): Zero-touch networking and service management / ZSM discussions; links to NFV / orchestration era
  • 2020: “Language Models are Few-Shot Learners” → LLM mainstream
  • Late 2022: ChatGPT public launch, rapid adoption (100M users in 2 months)
  • 2022: 3GPP intent-driven management/orchestration specifications
  • 2024: Anthropic MCP introduced for tool/context interoperability
  • 2025: Google A2A (agent-to-agent) and “year of agents” as interoperability becomes feasible
  • 2026: agentic AI appearing in production; agentic AI considered a 6G use case (March 2026 mentioned)

3) TM Forum autonomy ladder (L0–L5) mapped to telco concepts

  • L0: manual
  • L2: domain orchestration
    • Maps to orchestrators like NFVO / RAN domain orchestrator
    • Scripted automation/RPA-like behavior
  • L3: multi-domain orchestration and end-to-end service orchestration
    • Includes RAG use cases and approval gates
  • L4: agentic AI + multi-agent systems → minimal human intervention
  • L5: full autonomy
    • Human “informed,” not “in the loop”

Telco use cases highlighted

1) Autonomous fault management / autonomous network operations

Described as:

  • Detect → perceive context → reason/correlate signals → find root cause across a complex stack → propose plan → act → verify.

Complexity emphasized:

  • Fault localization spans:
    • passive elements
    • hardware/software/middleware
    • storage, networking
    • data fabrics, GPUs, AI platforms, etc.

A key pain point:

  • RCA is traditionally difficult

2) Proactive run-and-energy optimization (using predictive reasoning)

Example:

  • When low demand occurs in a geographic area/season, reduce energy by idling/high-power site tuning.

Key points:

  • Reasoning includes prediction (explicitly noted: not always LLM-based; predictive ML can drive decisions)
  • Guardrails:
    • Must not violate SLA / SLO
    • Must verify quality of experience
    • Roll back if results are not acceptable

3) “Dark Knock / Lights-out” network operations center

Defined as a TM Forum-labeled concept where a network ops center can:

  • detect, diagnose, remediate, verify
  • with no operator paging

Presented as a trending direction toward autonomous network operations, with multiple vendor-aligned initiatives referenced (e.g., Darktrace, Huawei ADN, Cisco intent-based operations, etc.).

Two key enabling concepts for production systems

1) Skills (procedural knowledge)

  • Skills are packaged procedural workflows (described like YAML artifacts, similar to VNF packaging concepts).
  • Problem solved:
    • LLMs may not know domain-specific operational procedures
    • If actions are random, systems can be messed up

Progressive disclosure behavior:

  • The LLM first sees only name/description
  • The agent loads the full skill after selection
    • reduces context overload and hallucination risk

2) Interoperability between agents and tools

  • A2A (agent-to-agent): agents communicate directly
  • MCP (Model Context Protocol): agents connect to external tools/data (monitoring systems, KPIs, inventories, repos)

Intermediary/orchestration nuance:

In telco, multi-domain interactions often require an orchestrator layer to facilitate coordination, so direct agent-to-agent isn’t always sufficient by itself.

Takeaways (as stated)

  • Agentic AI ≠ GenAI
    • Agentic AI executes actions toward goals, driven by a loop (perceive–reason–act–observe).
  • LLM is the reasoning engine, while tools/RAG/runtime/guardrails execute and validate.
  • Context quality is critical (“garbage in, garbage out”): the agent must build accurate context.
  • Guardrails + approvals exist (gated actions/HITL).
  • Dark Knock is actively pursued; autonomy targets (L4 → L5) are a major telecom direction.
  • The webinar claims agentic AI is no longer just roadmap:
    • deployments/demonstrations are happening
    • expected to grow through the 6G era

Main speakers/sources mentioned

  • Speaker: Karim Rabia (Red Hat)
  • Standards / industry sources: TM Forum, 3GPP, 6G (6G use case mention)
  • Protocols / tech sources: Anthropic MCP, Google A2A, Linux Foundation (handover noted)
  • Research papers:
    • “Attention Is All You Need” (2017)
    • “Language Models are Few-Shot Learners” (2020)
  • Market / industry source: NVIDIA State of AI (2024–2026 releases mentioned)

Original video