Production evidence for the age of AI-written code.
Engineering
We replayed four synthetic incidents through the SLO evaluator Belvador runs every minute. A plain threshold alert on the same traffic fired 16 times for one slow leak.
Andrew Babo6 min read
Product
Belvador does not have an MCP server. You can still give a coding agent what it needs: the issue's evidence, docs it can read as Markdown, and signed alerts your own code can turn into agent tasks.
Andrew Babo4 min read
Engineering
A route loader that throws during server rendering reaches the browser as a message and nothing else. We traced one error through three setups and counted what an error tracker received each time.
Andrew Babo5 min read
Guides
Error boundaries catch fewer errors than most people expect. We threw errors in eight places in a React 19 production build and recorded what the boundary, the root callbacks and the window each saw.
Andrew Babo4 min read
Engineering
We ran Belvador's scrubber on 16 synthetic error samples, fixed the three gaps that were ours and ran it again. Secrets with a known shape went; phones and a name still got through.
Andrew Babo5 min read
Product
We counted what a self-hosted Sentry 26.10.0 install pulls and starts, and set it next to a managed node in your own Hetzner project. Both keep data on your servers; the work is not the same.
Andrew Babo4 min read
Guides
A two-person team does not need a metrics platform on day one. It needs to know when users hit errors, when a job stops, when the site is down, and later, whether users are having a bad week.
Andrew Babo5 min read
Guides
We threw errors in eight places in a React 19 page and five places in a Node.js server, three runs each. Some crashed the server, some left a button that did nothing, and one hung a request forever.
Andrew Babo6 min read
Product
We put Sentry's browser SDK and Belvador's on the same page and recorded what each one sent. For an app an AI agent wrote, defaults, data location and proof of delivery matter before price.
Andrew Babo5 min read
Engineering
We ran ten common ways of pinging a heartbeat from a shell job in bash, dash and the /bin/sh Ubuntu 24.04 ships. Three reported success for a failed job, one never ran the job at all.
Andrew Babo5 min read
Engineering
We pushed one synthetic day of 2,268 errors through four alert policies. One alert per event sent 2,268 messages; the policy most teams end up with sent 5, and two of those were still noise.
Andrew Babo5 min read
Guides
Keeping error data in the EU says nothing about where an AI model processes the error you paste into it. We compared four AI providers and found storage and processing are separate dials.
Andrew Babo4 min read
