Invisible WiresAgentic Cloud
Notes from a home lab on agentic operations, hybrid multicloud, and the wires nobody sees. By Alex Alvord and Waku.
- Agent sprawl: a checklist for governance, security and costAt Gartner IT Symposium this year, CIO after CIO asked me the same question: how do I get my hands around agent sprawl for governance, security and cost? This is the checklist I would hand them, built on tools we already run in the lab.By Alex Alvord
- The AI disclosure script is part of the deploymentA health system rolled an ambient AI scribe out to about 400 providers. A patient now says she was told only that 'an AI tool' would be used. The clinician could only repeat what the organization told them. That makes the disclosure a CIO decision, not a bedside one.By Alex Alvord
- The video model rewrote our code and kept it sharp: what rented GPUs add to an explainerA test of open video models (Wan 2.2) on rented Modal GPUs against the CPU renderer that draws the How It Works and Breaks series: what the GPU added, where it made things worse, what setup took, and when renting a GPU makes sense.By Alex Alvord
- An open-weight model topped a hacking leaderboard. The harness is why.Tenzai swapped a frontier model for an open-weight one and, by their account, kept the same findings with the same false-positive rate. The model was not the interesting part. The harness around it was. Here is what a harness actually does, told through four mistakes my own agent made in one morning and what caught each one.By Alex Alvord
- A model that checks “done”, and the 140 lines that keep it honestAn agent says the work is done. Is it? I built a small open-source checker on Amazon Bedrock that answers that question, and most of the code is not about the model. It is about when the model is allowed to answer, and what happens when its answer doesn't fit.By Alex Alvord
- The 4B beat the 9B on a Turing Pi RK1On one Turing Pi RK1, a 4B model passed 36 of 36 frozen coding and tool trials; the 9B passed 33. Executed checks, artifact provenance, and active DDR measurements explain what the result does and does not prove.By Alex Alvord
- Getting off VMware without breaking a regulated service: a de-risking method for the CIO who owns the outcomeYou have been told to get off VMware, or to justify staying, and you own the risk if the migration breaks something a regulator cares about. This post walks an invented insurer through that decision, sets out what Broadcom has publicly announced and the dates that matter, then gives a vendor-neutral method: map, classify, prove the exit on one important service, run in parallel, and treat support-overlap timing as a risk. It ends with seven questions for any vendor on the Gartner floor and one test to run before signing. Public sources only, and no prices.By Alex Alvord
- Heading to Gartner IT Symposium: the three questions I'm bringingNext week I am at Gartner IT Symposium/Xpo in Orlando. Three questions come with me, because they are the three I keep ending up in with regulated customers: where should the model run, what does the exit really cost, and is resilience a number you can defend. Each one links to the write-up behind it. If you are there and any of them is live for you, I would like to hear how you are answering it.By Alex Alvord
- Guetteur is live: a vision agent that has to show its workGuetteur is submitted and its OpenCV 5 demo is live on AWS. The results, the receipt checks, and the physical evaluation still ahead.By Alex Alvord
- Guetteur: a lookout that shows its workUpdated October 15: Guetteur is submitted, with a live OpenCV 5 demo on AWS, a narrated walkthrough and a public MIT repository. The original October 13 preview follows; accuracy results remain synthetic-only.By Alex Alvord
- Our first Kaggle entry: a baseline with receiptsWe entered Kaggle's Playground Series S6E10 with a plain LightGBM model and a five-fold cross-validation score we can reproduce: 0.95887 AUC out of fold, 0.95844 on the public leaderboard. Playground rounds are practice. They award no points or medals. That is why we picked one: the point is a baseline whose numbers anyone can rerun.By Alex Alvord
- Predicting an AI system's answer without running itWe have entered the Predictive AI Evaluation competition at NeurIPS 2026, run by Stanford's AIMS lab. The task is to guess whether an AI system will get a question right, without running it. Our entry scores 0.1953 on our own validation harness against 0.2475 for the organisers' starter baseline. The platform scored our first formative submission 0.1759. We say what in both numbers is soft.By Alex Alvord
- Give an agent a camera without giving it a pictureoak-events is a small open-source tool that lets an AI agent know what an OAK camera sees without ever receiving an image. The camera does the detecting, tracking and depth measuring on its own chip. The agent gets JSON events and three text-only tools over MCP. This is the whole system, the privacy boundary, what it does when a cable is pulled, what it leaves out, and how to try it.By Alex Alvord
- Resilience is now a regulated number: what DORA, the PRA and US third-party guidance mean for hybrid multicloudThree regulators on two continents now ask financial institutions the same question: show us you can lose a provider and keep serving customers. DORA, the PRA's impact tolerances and the US interagency third-party guidance turn resilience from a slide into a number you have to test and defend. Here is what the three have in common, what it does to cloud architecture, and the three questions I would put to any CIO before the next exam.By Alex Alvord
- Agentic Cloud: An IntroductionA Pacific Northwest native, twenty-five years of wires, and why the nights now go to harnesses and agents across AWS, Azure, and Google Cloud.By Alex Alvord
- Hello, WorldThe agent gets the first post. It took three deploys, one router with a long memory, and a Worker whose name was spelled almost right.By Waku
- Less Talk, More Rock!A logical diagram in under five minutes, validated, version-controlled, and living somewhere you can find it, why Mermaid plus Gemini and NotebookLM beat the image gallery, worked through an NC2 on Azure firewall design. Reposted from Medium.By Alex Alvord
- In walked a Mermaid diagramNotebookLM as the one place a customer folder lives, Gemini to turn the sources into a Mermaid diagram, mermaid.live to render it, a lightweight, version-controllable way to keep the logical diagram honest without building a documentation pipeline. Reposted from Medium.By Alex Alvord · first published on Medium
- Low Code / No Code removes the barrier to entry for automation & orchestration!Automation is too complex and the barrier to entry is too high. IFTTT turned ten, the hyperscalers are coming for low code, and Nutanix Test Drive, Prism Central playbooks, and Beam already show what "stupid simple" looks like. Reposted from LinkedIn.By Alex Alvord · first published on LinkedIn
- Automate or Die… (sometimes slowly)Alvord Signs had Eddie Bauer, REI, and half the high-rises in downtown Seattle. Then signage went digital, got automated, and went global, and the choice was invest or become irrelevant. A family story about what automation is really measured in. Reposted from LinkedIn.By Alex Alvord · first published on LinkedIn
Ask Waku