Production evidence for the age of AI-written code.
Guides
Cursor's Debug Mode is built for bugs you can reproduce. A production error already happened to someone else, so first turn it into steps. A stack with the input did that; a complaint did not.
Andrew Babo5 min read
Product
Changing a DSN is the easy part. We ran the official Sentry Node.js SDK with two configs and listed every item it sent, and what a Sentry-compatible backend does with each one.
Andrew Babo5 min read
Guides
Open the published page, open the console, reload. In our tests the first red line told five kinds of blank or frozen page apart, and each one has a different fix.
Andrew Babo5 min read
Product
Seer reads your Sentry data and linked code inside Sentry. A coding agent reads what you hand it in your repo. Either way the risk sits at the handover: a wrong analysis the next step believes.
Andrew Babo4 min read
Engineering
Ten AI fixes in our test were wrong, and each came with a test that failed before the fix and passed after. The tests proved something, not the bug customers hit. What to require instead.
Andrew Babo4 min read
Engineering
We ran the official OpenAI instrumentation for Python in five settings. By default it uses older attribute names, records nothing for the Responses API and leaves agent and tool spans to you.
Andrew Babo3 min read
Guides
Production errors are not in Claude Code's context until you put them there. From a bug report alone, a Claude model fixed the wrong bug 8 times in 8; from an error packet, the right one 8 in 8.
Andrew Babo5 min read
Guides
An agent with your whole repo still guesses from a complaint. Three models, two bugs, five handovers, 108 answers: a bug report found the right bug 8 times in 24, a short error packet 24 in 24.
Andrew Babo6 min read
Engineering
Reporting an error and then rethrowing it did not double it in our lab. Reporting it and then throwing a wrapper did. Six patterns, counted with the Sentry browser SDK and our own.
Andrew Babo4 min read
Engineering
We retried a failing payment job with the Sentry Node SDK and counted every event. Ten failed jobs sent 30 events; one that recovered still sent 2. Dedupe did not help.
Andrew Babo6 min read
Product
Sentry's MCP server lets an agent search and fetch issues itself. A pasted packet hands it one issue, ready to read. We measured what each costs an agent, and where each one fits.
Andrew Babo5 min read
