This article argues that no software system can be guaranteed to never fail, since failures can arise from factors beyond code quality. It explains how to design for failure by predefining system behavior, reducing impact, and measuring recovery, covering topics from fault tolerance to resilience strategies.
Reliability in AI systems depends on the whole decision pipeline—data, features, model, policy, action, and feedback—not just the model itself. As AI moves toward agentic action in live operations, systems must be designed to withstand changes and prevent silent failures.
The book describes how Google combines software engineering and operations to manage reliable production systems. It covers practices such as automation, service level objectives, error budgets, and blameless postmortems.
The 2027 ICRE conference in Harbin, China, July 16–18, sponsored by multiple institutes and IEEE, focuses on AI-driven reliability and sustainable, resilient systems, with four tracks and publication in major indexes.
Translated title. Original: NerveCenter: agente IA con memoria para SRE
NerveCenter is a memory-enabled SRE AI agent that uses past incident logs to generate precise, context-aware commands instead of generic suggestions. Built as a four-person proof-of-concept (sync4ce) with a Streamlit UI, it demonstrates how referencing post-mortems can improve incident response by tying commands to documented outcomes.
Open Oncall is an open-source on-call workflow for Agent Swarm that automates incident handling from Slack alerts to one-click approvals and PRs. It supports multi-provider failover, multiple alert sources, and a documented set of scripts, apps, and workflows that users configure and import to run the entire process.
The document explains how small faults in shared cloud infrastructure can spread across services such as Azure, Microsoft 365, and Teams. It outlines common causes of outages and how understanding them can help IT teams plan for disruptions.
Meta’s incident response team built an AI teammate to connect alerts, deployments, and past incidents, then propose actions for human approval. In a 23-item trial, three items needed direct human attention, while 18 went through normal review and CI.
Translated title. Original: Automação não é suficiente: a autonomia em TI precisa saber quando parar | TI INSIDE Online
Autonomy in IT must be bounded by data quality, context, and governance; automation handles routine tasks, while autonomous systems must decide when and how to act within defined guardrails to avoid costly errors.
A developer explains that MCP does not define retry behavior; retries are handled by the harness within the host application. The article details MCP roles, protocol components, and how recent spec changes shift state management and caching responsibilities to the harness and host.