Numtide built a production Kubernetes operator for Multigres using a design‑first, AI‑assisted workflow. They treated the user‑facing spec as the immutable contract and turned AI into a factory of narrow, orchestrated agents, fixing skills when they misbehave. The result proves AI can handle complex operator code while humans steer architecture and safety.
AWS adds a custom Model Context Protocol (MCP) server that lets its DevOps Agent pull from 20+ node‑level logs and metrics on EKS workers. The agent can now autonomously pinpoint issues like CrashLoopBackOff without manual SSH, speeding root‑cause analysis and reducing downtime.
Google’s SRE team is moving beyond scripted automation to agentic AI that can investigate incidents, suggest fixes, and influence design decisions across the SDLC. The shift promises faster root‑cause analysis, reduced human toil, and more reliable services as AI becomes a force‑multiplier rather than a replacement.
Coinbase’s post‑mortem reveals that a cooling‑unit malfunction in an AWS us‑east‑1 data hall knocked out EC2 instances and its matching engine, shutting trading for eight hours. The incident highlights how reliance on a single availability zone can cascade into platform‑wide outages and why multi‑zone resilience is essential.
AWS has launched the FinOps Agent in public preview, an AI‑driven assistant that detects cost anomalies, pinpoints the change that caused them, and drafts remediation tickets in Jira or Slack. It lets teams ask natural‑language billing questions and automates routine reporting, tightening the feedback loop between engineering and finance.
The article warns that prompt-to-app tools deploy code on the AI vendor’s cloud, preventing monitoring, testing, compliance, and multi‑cloud strategies. This creates vendor lock‑in that breaks CI/CD, security audits, and operational control, forcing teams to duplicate environments or abandon AI‑generated prototypes.
Dropbox built a system that pulls threat‑model docs into pull‑request reviews using the Model Context Protocol, large language models, and its internal Dash AI. It flags code changes that stray from the documented security requirements, closing the design‑to‑code gap and reducing manual oversight.
Tenet Security showed that a single forged Sentry error, sent through a public DSN, can trick AI coding assistants like Claude Code and Cursor into executing attacker‑controlled code on a developer’s machine. The attack bypasses firewalls, EDR and user prompts, exposing credentials and repos at scale across thousands of organizations. Mitigation requires tightening trusted data handling in AI agent runtimes.
You’re caught up.
That was all 8 stories in the 16 June issue.