arxiv: 2604.03512 · v2 · submitted 2026-04-03 · 💻 cs.AI

Recognition: no theorem link

ActionNex: A Virtual Outage Manager for Cloud Computing

Zhenfeng Lin , Haoji Hu , Ming Hao , Xuchao Zhang , Ryan Zhang , Junhao Li , Ze Li , Oleg Kulygin

show 6 more authors

Chetan Bansal Hatay Tuna Murali Chintalapati Sheila Jiang Salman Zafar Angie Anderson

Authors on Pith no claims yet

Pith reviewed 2026-05-13 19:15 UTC · model grok-4.3

classification 💻 cs.AI

keywords cloud outage managementagentic systemaction recommendationcritical eventshierarchical memoryproduction pilot

0 comments

The pith

ActionNex compresses multimodal cloud signals into critical events and matches them against hierarchical memory to recommend next actions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents ActionNex as an agentic system that supports end-to-end outage assistance in large-scale cloud operations. It ingests signals such as telemetry and human communications, compresses them into critical events that mark meaningful state transitions, and aligns those events with a memory structure containing distilled playbook knowledge, past outage episodes, and live context. A reasoning agent then generates role- and stage-conditioned action recommendations. If the approach holds, operators gain real-time guidance that reduces manual triage and coordination under partial observability, while executed actions feed back to refine the system over time.

Core claim

ActionNex ingests multimodal operational signals and compresses them into critical events representing meaningful state transitions. It couples this with a hierarchical memory subsystem of long-term Key-Condition-Action knowledge distilled from playbooks, episodic memory of prior outages, and working memory of the live context. A reasoning agent aligns current events to preconditions, retrieves relevant memories, and generates actionable recommendations, with executed human actions serving as implicit feedback for continual self-evolution. On eight real Azure outages totaling 8M tokens and 4,000 critical events, the system achieves 71.4% precision and 52.8-54.8% recall against two ground-tru

What carries the argument

The perception layer that compresses multimodal signals into critical events representing state transitions, coupled with a hierarchical memory subsystem (long-term KCA knowledge, episodic memory, and working memory) that a reasoning agent uses to align events and retrieve recommendations.

Load-bearing premise

Compressing multimodal signals into critical events accurately captures meaningful state transitions and the memory retrieval reliably produces useful recommendations without significant irrelevance.

What would settle it

A fresh outage in which the generated recommendations consistently diverge from the sequence of actions later taken by the responding operators.

Figures

Figures reproduced from arXiv: 2604.03512 by Angie Anderson, Chetan Bansal, Haoji Hu, Hatay Tuna, Junhao Li, Ming Hao, Murali Chintalapati, Oleg Kulygin, Ryan Zhang, Salman Zafar, Sheila Jiang, Xuchao Zhang, Ze Li, Zhenfeng Lin.

**Figure 1.** Figure 1: ActionNex Framework present ActionNex, a continual, self-evolving outage management solution which uses action as general abstraction for knowledge, decision-making and tasks. The core idea is to learn what actions to take and when to take the action from human experiences while operating in a human-agent hybrid system so that it can recommend actions in a timely fashion (and execute them in the future). … view at source ↗

**Figure 3.** Figure 3: Accuracy over Stages for G1 and G2 instance, emphasizing episodic-memory analogies in early stages to improve recall, and shifting toward stricter condition gating later to preserve precision. 6 Conclusions We presented a production-grade agentic system for outage management that combines multimodal context, criticalevents, and action recommendations with a hierarchical memory layer centered on long-term… view at source ↗

read the original abstract

Outage management in large-scale cloud operations remains heavily manual, requiring rapid triage, cross-team coordination, and experience-driven decisions under partial observability. We present \textbf{ActionNex}, a production-grade agentic system that supports end-to-end outage assistance, including real-time updates, knowledge distillation, and role- and stage-conditioned next-best action recommendations. ActionNex ingests multimodal operational signals (e.g., outage content, telemetry, and human communications) and compresses them into critical events that represent meaningful state transitions. It couples this perception layer with a hierarchical memory subsystem: long-term Key-Condition-Action (KCA) knowledge distilled from playbooks and historical executions, episodic memory of prior outages, and working memory of the live context. A reasoning agent aligns current critical events to preconditions, retrieves relevant memories, and generates actionable recommendations; executed human actions serve as an implicit feedback signal to enable continual self-evolution in a human-agent hybrid system. We evaluate ActionNex on eight real Azure outages (8M tokens, 4,000 critical events) using two complementary ground-truth action sets, achieving 71.4\% precision and 52.8-54.8\% recall. The system has been piloted in production and has received positive early feedback.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ActionNex gives a working agentic pipeline for cloud outage triage with real metrics from eight cases, but the evaluation is too thin on baselines and ground-truth details to judge whether the architecture actually improves on simpler retrieval.

read the letter

ActionNex turns multimodal outage signals into critical events and uses a three-layer memory system—distilled KCA knowledge from playbooks, episodic outage history, and live working memory—to suggest next actions. It reports 71.4% precision and 52-55% recall on eight real Azure outages and has run in a production pilot with positive feedback. That is the core contribution: a concrete, end-to-end system that closes the loop with implicit human feedback for ongoing updates.

Referee Report

2 major / 1 minor

Summary. The paper presents ActionNex, a production-grade agentic system for end-to-end outage assistance in cloud computing. It ingests multimodal signals, compresses them into critical events representing state transitions, and uses a hierarchical memory subsystem (long-term KCA knowledge, episodic memory, working memory) to generate role- and stage-conditioned next-best action recommendations. The system is evaluated on eight real Azure outages involving 8M tokens and 4000 critical events, achieving 71.4% precision and 52.8-54.8% recall against two ground-truth action sets, and has received positive feedback in a production pilot.

Significance. If the results hold, this work offers a meaningful advance in applying agentic AI to real-world operational challenges in large-scale cloud systems. The integration of perception via critical events with hierarchical memory retrieval, combined with implicit feedback for self-evolution, addresses practical needs in outage management. The use of real outages and production deployment provides strong evidence of applicability, though the limited scale of evaluation (eight cases) tempers the generalizability claims.

major comments (2)

[Evaluation section] The reported 71.4% precision and 52.8-54.8% recall on eight outages are presented without any baseline comparisons (e.g., random action selection, simple playbook lookup, or non-hierarchical retrieval). This omission makes it difficult to assess whether the hierarchical KCA/episodic/working memory alignment contributes meaningfully beyond simpler methods, which is load-bearing for the central performance claim.
[Evaluation section] Details on the construction of the two complementary ground-truth action sets are insufficient. It is unclear how these sets were built independently of the distilled KCA knowledge and whether they incorporate validation steps to avoid selection bias or circularity with the system's memory retrieval, undermining verification of the precision/recall metrics.

minor comments (1)

[Abstract] The abstract mentions 'positive early feedback' from the production pilot but provides no quantitative details or specific metrics on user satisfaction or impact, which would strengthen the deployment claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the evaluation section. We address each major point below and will revise the manuscript accordingly to strengthen the claims.

read point-by-point responses

Referee: [Evaluation section] The reported 71.4% precision and 52.8-54.8% recall on eight outages are presented without any baseline comparisons (e.g., random action selection, simple playbook lookup, or non-hierarchical retrieval). This omission makes it difficult to assess whether the hierarchical KCA/episodic/working memory alignment contributes meaningfully beyond simpler methods, which is load-bearing for the central performance claim.

Authors: We agree that baseline comparisons are essential to isolate the contribution of the hierarchical memory alignment. In the revised manuscript, we will add results for random action selection, a simple playbook lookup baseline, and a non-hierarchical retrieval variant, all evaluated on the same eight Azure outages. This will provide direct context for the reported precision and recall figures. revision: yes
Referee: [Evaluation section] Details on the construction of the two complementary ground-truth action sets are insufficient. It is unclear how these sets were built independently of the distilled KCA knowledge and whether they incorporate validation steps to avoid selection bias or circularity with the system's memory retrieval, undermining verification of the precision/recall metrics.

Authors: We acknowledge the lack of detail on ground-truth construction. The revised manuscript will expand this section with a full description of the independent expert annotation process used to create both sets, explicitly noting their separation from KCA distillation and the validation steps applied to mitigate bias and circularity. revision: yes

Circularity Check

0 steps flagged

No significant circularity: evaluation grounded in independent real-world outages and ground-truth sets

full rationale

The paper presents an engineering system description rather than a mathematical derivation chain. ActionNex compresses multimodal signals into critical events and uses hierarchical memory (KCA distilled from playbooks/historical data, episodic, and working memory) to generate recommendations via alignment and retrieval. The load-bearing performance claims rest on evaluation against eight real Azure outages using two complementary ground-truth action sets, with implicit human feedback for self-evolution. No equations, fitted parameters renamed as predictions, or self-citation chains reduce the reported precision/recall to inputs by construction. The results are externally grounded in production data and human actions, rendering the architecture self-contained without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The system rests on domain assumptions about signal compression and feedback utility, with new entities like critical events and KCA knowledge postulated without independent evidence beyond the reported metrics.

axioms (2)

domain assumption Multimodal operational signals can be reliably compressed into critical events representing meaningful state transitions.
Central to the perception layer described in the abstract.
domain assumption Human actions provide valid implicit feedback for continual self-evolution of the system.
Used to enable learning in the human-agent hybrid setup.

invented entities (2)

Critical events no independent evidence
purpose: Represent meaningful state transitions from multimodal signals
New abstraction introduced for the system.
Key-Condition-Action (KCA) knowledge no independent evidence
purpose: Long-term knowledge distilled from playbooks and historical executions
Invented structure for memory subsystem.

pith-pipeline@v0.9.0 · 5571 in / 1464 out tokens · 63864 ms · 2026-05-13T19:15:25.016927+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 6 internal anchors

[1]

Anonymous. 2026. REMem: Reasoning with Episodic Memory in Language Agents. InICLR

work page 2026
[2]

Abrar et al. Anwar. 2024. ReMEmbR: Long-Horizon Spatio-Temporal Memory for Robot Navigation.arXiv:2409.13682(2024)

work page arXiv 2024
[3]

Zouying et al. Cao. 2025. Remember Me, Refine Me: Dynamic Proce- dural Memory for Agent Evolution.arXiv:2512.10696(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[4]

Ruowei et al. Fu. 2025. OncallX: LLM-Powered Multi-Agent Collabo- ration for On-Call Automation. InASE

work page 2025
[5]

Hamadanian

Pouya et al. Hamadanian. 2023. A Holistic View of AI-Driven Network Incident Management. InHotNets

work page 2023
[6]

Bowen et al. Jin. 2025. Search-R1: Training LLMs to Reason with Search via Reinforcement Learning.arXiv:2503.09516(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[7]

Jian-Guang et al. Lou. 2013. Software Analytics for Incident Manage- ment of Online Services. InASE

work page 2013
[8]

Jiacheng et al. Mao. 2025. Agentic Troubleshooting Guide Automation for Incident Management.arXiv:2510.10074(2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[9]

Siru et al. Ouyang. 2025. ReasoningBank: Scaling Agent Self-Evolution with Reasoning Memory.arXiv:2509.25140(2025)

work page internal anchor Pith review arXiv 2025
[10]

Timo et al. Schick. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools.arXiv:2302.04761(2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[11]

Chenxu et al. Wang. 2025. Towards LLM-Based Failure Localization in Production Networks. InSIGCOMM

work page 2025
[12]

Zefan et al. Wang. 2024. RCAgent: Cloud Root Cause Analysis with Autonomous Agents. InCIKM

work page 2024
[13]

Tianxin et al. Wei. 2025. Evo-Memory: Benchmarking Test-Time Learning with Self-Evolving Memory.arXiv:2511.20857(2025). 6 ActionNex: A Virtual Outage Manager for Cloud Computing

work page internal anchor Pith review arXiv 2025
[14]

Tianxin et al. Wei. 2026. Agentic Reasoning for Large Language Models. arXiv:2601.12538(2026)

work page arXiv 2026
[15]

Menglin et al. Xia. 2026. Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity.arXiv:2602.03315(2026)

work page arXiv 2026
[16]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations

work page 2022
[17]

Zhaoyang et al. Yu. 2025. Triangle: Empowering Incident Triage with Multi-Agent Systems. InASE

work page 2025
[18]

Aohan et al. Zeng. 2023. AgentTuning: Enabling Generalized Agent Abilities for LLMs.arXiv:2310.12823(2023)

work page arXiv 2023
[19]

Wentao et al. Zhang. 2025. AgentOrchestra: Orchestrating Multi-Agent Intelligence.arXiv:2506.12508(2025). 7

work page arXiv 2025