Operations | Monitoring | ITSM | DevOps | Cloud

I have an Upsun project that's nothing but proofs of concept. It's a dashboard, basically. Each POC gets its own tile. Click in, and you land on a page with three tabs. The first tab is a written explanation of what the POC argues. The second tab is the POC itself, with a built-in demo that automates a walkthrough of the feature so the recipient can watch it run without me on the call.

This post covers some techniques to get agents to validate their work. We’ll cover why this is important, a variety of types of checks you can employ, and mechanisms you can use to enforce them.

AI coding tools promise faster development. What they don't show you is the queue forming at the pipeline, the security scanner you bypassed to stay fast, or the cost dashboard with a line now labeled "unknown" that is steadily growing. In May, we shipped 60+ features in 31 days across the entire delivery system: not just the editor, but everything downstream of it.

You don’t have an SRE. There’s no platform team. Your “monitoring strategy” is someone checking Slack for error alerts. When production breaks, the same two or three senior devs drop everything to debug. Sound familiar? Most APM tools are built for organizations with dedicated operations staff. They assume someone has time to configure dashboards, tune alert thresholds, and learn a complex query language. That person does not exist on your team.

You deploy on Friday. Sidekiq starts failing on a job that worked fine in staging. Your error tool shows you a NoMethodError on line 47. But it doesn’t tell you that the job only fails when processing records created after the migration you ran on Thursday. The stack trace is correct and completely useless at the same time. This is the core problem with general-purpose error monitoring on Rails apps. Rails teams deal with N+1 queries that cascade into timeout errors.

DNS monitoring should be simple. You want to know if something changed. You want to know if a record propagated. You want to know if a phishing site just went live with your brand name in the domain. But in practice it takes work. You log in to a dashboard. You click through menus. You run a check, copy the output, paste it somewhere else. You repeat that process every time someone on the team asks a question. AI assistants like Claude and ChatGPT could help.

In December, we shared our plans to introduce pricing for self-hosted runners. You told us loud and clear that a free option matters. Today, as Premium Runners become generally available, we are happy to share that we will continue to have a free tier, which includes the use of up to 100 self-hosted runners as part of your plan. If your team needs more scale, dedicated support, or advanced management features, you can upgrade to Premium Runners when you’re ready.

Every quarter, the same scene plays out in boardrooms across the Fortune 500. The CEO asks: “What is the return on everything the company is spending on AI?” The CTO talks about productivity gains and developer velocity. The CFO points at a cloud bill that doubled but cannot isolate which line items are AI. The board nods politely and tables the discussion until next quarter, when the same question will produce the same non-answer. (If this sounds familiar, you are not alone. Keep reading.)

CloudZero originated as a way to make sense of your cloud costs. Costs spread across bills with billions of line items belonging to resources that might or might not have been tagged (or taggable), spun up by engineers working across teams, on different microservices, features, and products, that served a wide range of customers. Kubernetes. Multi-cloud. Check, check, check.

For many development teams, a load test starts with a set of assumptions. You pick 100 virtual users because it sounds reasonable. You ramp for 30 seconds because that's what the tutorial showed. You set a 500ms threshold because it feels like a good target. The test passes, you ship the release, and production falls over at 6 p.m. on a Tuesday because your synthetic load never resembled how real users interact with your application.