In public
Skoor, skoored.
We run agents unattended every night — a build node that reviews production, opens pull requests, and scores itself while nobody watches. We hold ourselves to the same gates we tell you to hold yours to, and we publish them live: our own, unedited, gates and all.
Our gates, right now
Read live from the same status our overnight pipeline computes. Most read UNMET or UNMEASURED — that is the honest state of a company still earning them, and a number we have not measured is exactly the lie this whole company exists to refuse.
Source: api.skoor.ai/api/v1/gstack/status— the same status our overnight pipeline computes. Published unedited, gates and all.
Four times it looked green while it was broken
These are real, from our own machine, inside a single stretch of work. We publish them because they are the argument. Every one is the same shape: a system that looked fine because the thing meant to check it was not running.
The trust metric that could not go red
What happened
Our golden-path check scored three of five legs. The two it left out — can a customer log in, can they get value — were the two that mattered. It reported GREEN for 27 consecutive days.
Why it looked green
Every dashboard said healthy. The streak was cited as proof the product worked.
What caught it
A read of the scoring code, not the dashboard. Fixed so every leg counts; it went RED the next run, correctly, and stays RED until two credentials are provisioned.
Ninety CI runs that never started
What happened
For three days, every continuous-integration run on one repo failed at startup — a platform outage, not a code fault. Zero of ninety runs completed.
Why it looked green
No red X on any pull request, because the checks never ran to fail. Absence of failure read as success.
What caught it
A required-status-check made the silence visible: pull requests suddenly could not merge. The outage had been invisible precisely because nothing was watching whether the checks ran.
The gate that was dead at step one
What happened
Our pre-merge verification script called a runtime flag that a newer runtime had removed. It exited at its first line, every time.
Why it looked green
Every pull request claimed 'verify passed' — because nobody re-ran it after the runtime updated. A gate that does not run looks identical to a gate that passes.
What caught it
Running it by hand on a clean checkout. One flag removed; the gate came back — and immediately caught two real regressions in the same week, including two of our own.
Eleven pages shipped unreachable
What happened
A set of pages went live with no navigation link, no homepage link, and no sitemap entry. Findable only by typing the URL.
Why it looked green
Every build was green. An unreachable page compiles perfectly; nothing in the build can tell you a page has no way in.
What caught it
Reading the routes, not the build log. Wired in, plus a rule: a change that adds a route adds its link and its sitemap entry in the same change.
Why we publish the failures
A vendor that only shows you its green streak is showing you the exact thing that lied to us for 27 days. The failures are the proof, and they are unfakeable — no competitor can publish our incident history, because it is ours. This is what it looks like to run agents unattended honestly: not never breaking, but catching it, and saying so.
It is also the whole product in one lesson. Automation that is not measured fails silently and looks identical to success. The answer is not to automate less — it is to measure the automation, and measure the measurement. That is what an Agent Skoor is.
We keep our own board. Start keeping yours.
The assessment produces your baseline — and you can start the roster yourself, free, in two minutes.