Amp

Try it on your own code: https://sourcegraph.com/deep-search?utm_source=youtube-incident-response&utm_medium=social A cart page returns 500. One line in the frontend log: "could not retrieve cart." Nothing about why. So you grep the error string and get two hits, Redis and Spanner. It looks like the whole answer. It isn't. A third backend throws the same failure in different words, and grep has no way to know it exists. Ship the two-file fix and the alert comes back. This is 85 seconds of triage on an open-source codebase: the page, the log line, the grep that half works, the question asked in plain English, and the same alert closing green nine minutes later. Why it matters depends on your seat. If you run engineering, the cost was never the nine minutes. It's the repeat incident, and the fact that one senior engineer was the only person who knew the third backend was there. If you own platform or developer experience, this is the gap your search tooling leaves open. Grep finds the string you typed, not the code that says it differently, and that gap widens with every service, every repo, every acquisition. If you own risk, swap "incident" for "CVE" and it's the question you ask after every advisory: did we find every place this pattern exists, or only the places that phrase it the way we guessed? Chapters: 0:00 The page 0:15 What the log actually says 0:26 The grep that finds 2 of 3 0:37 Asking instead of guessing 0:49 All three, with citations 1:08 Same page, different ending Production notes: The codebase is real and unmodified: GoogleCloudPlatform/microservices-demo (Apache-2.0) at commit b9a978d. Those three files genuinely throw the same error three different ways, and a literal grep genuinely misses one. The Slack alert is a re-creation.