Manufacturing Alerts To ReActive Agents: Self-Evolving AI Agents on K8S for Predictive…
summary
Claims a production pilot cuts diagnosis from days to under an hour with ~90% field acceptance, using an anomaly score (example 0.92) published to Kafka topic anomaly.events and a Kubernetes cloud back end. Architecture centers on a ReAct core with modular skills loaded at invocation, a leader/researcher/critic multi-agent deep search that spawns isolated jobs, and a self‑evolving pipeline that promotes skills after usage thresholds (≈3/week candidate, ≈5/week promotion). Uses Argo CD for declarative skill rollout and favors job-per-query researcher pods over warm pools for cost and isolation.
AI-written summary · about
chapters (13)
- 0:00Self-evolving AI agents on K8S
- 1:03Threshold alerts in manufacturing
- 3:08Bearing vibration event flow
- 5:20Version 0.1 to modular agents
- 7:06Domain-agnostic agent core
- 9:59Skill layers and types
- 11:00Skills as K8S data with Argo CD
- 12:48Self-evolving agents
- 15:38Skill evolution methods
- 17:52Multi-agent deep search
- 21:27Job-per-curry vs warm pool
- 23:12Edge-to-report event processing
- 24:39Key takeaways
related talks
33mStrategic AI Agents for K8s SRE: Beyond Ad-hoc Prompting
30mAI Agents for Incident Management
30mArchitecting Integrity: Responsible Software Evolution in the AI Era
18mFrom Reactive to Resilient: Agentic AI
19mNo Token Budget: Agentic Software Factories on Kubernetes
07h 37mAI Agents Conference 2026 - Day 3
42mAI for security - friend or foe?
19mAI Agents for Incidents: The Good, the Bad, and the Ugly