Episodes
Two hosts, a few minutes, the week's AI news for the people who have to operate it. Sources are real and attributed; the takes are about what it means for the systems you run.
AI Infrastructure Hits the KEV List, and Agents Discover the Cache
- Three of seven KEV additions were AI/ML infrastructure — LiteLLM (CVE-2026-59822): a fabricated Authorization header on the MCP endpoint fell through to an OAuth2 passthrough that built an empty auth object. Fixed in 1.84.0. Kestra (CVE-2026-49869): OS command injection in the orchestrator running your pipelines.
- Agents burn ~5x the tokens — 85% of it cached. a16z on OpenRouter data: ~7.3T tokens/7-day average from agents vs ~1.5T from humans, agentic usage up 14x since February. Your token count and your bill have decoupled; cache hit rate is what sits between them.
- Claude Fable 5.1 GA — $10/M in, $50/M out, cache reads cut 75% to $0.25/M. Agentic economics are being re-priced around the cache: prompt stability is now a cost control.
Homework: LiteLLM to 1.84.0 with real auth in front of every MCP endpoint; patch Kestra and give every AI service an owner; measure cache hit rate as a first-class metric; re-run your cost model — our calculator is free.
Free tools from this show: the calculators and the MCP server · get the field note by email
The Week AI Became an Ops Problem: Agent Reliability, Ox Alpha & Stateless MCP
- The agent reliability reckoning — Cisco: 85% piloting agents, only 5% in production. Reliability, not capability, is the blocker; the "rebuild era" is about durable execution, state, observability and recovery.
- Z.ai's "Ox Alpha" — an open-weight reasoning model topping leaderboards, weights out this week. Open + agentic + self-hostable changes the build-vs-buy math.
- Inherent's Faraday — a 27B-parameter agent beat frontier systems at reproducing scientific papers. The engineering around the model matters as much as the model.
- MCP goes stateless — sessions dropped, OAuth 2.1 auth added — run MCP servers behind a standard load balancer on Kubernetes.
Homework: build the boring reliability layer, keep an eval harness warm for open models, judge the workflow not the parameter count, and read the stateless MCP spec before you scale.
Free tools from this show: the calculators and the MCP server · get the field note by email
More from AIOps Field Notes
Clear, no-hype explainers on the AI actually reshaping ops — watch the video series on YouTube, and catch our sibling shows on networking and security.
Watch on YouTube →Get the field note by email
A short email when the calculators get new numbers, the directory gets new entries, and the week's episode lands. No hype — unsubscribe in one click.
Your address is used to send the field note and nothing else.
Who builds these tools
AIOps Field Notes is written by James Son, a network, security, and automation engineer, and everything here is tested in a live multi-vendor lab. That same lab work is behind two paid things for teams who run the infrastructure underneath the AI: