On-call
Nobody opens a log without the assistant
The rotation still pages. The first ten minutes of every incident are a prompt. You find out who can still read a stack trace when the model is down — or you find out during the outage.
Practice. Fly. Fix.
See who can still debug with the assistant off.
Short drills on real broken code. Incident simulations on a live host. A report that names who is drifting — before an outage does.
No card. No sales call.
Lead skill 25 mm/s Unaided
The evidence
Developers using AI assistants scored 17% lower on code comprehension. That’s Anthropic’s own RCT. Kaerus measures the skill that number is about.
How it works
Three rooms. The assistant stays closed.
-
01 — Drill
Fix code that is actually wrong
A broken function, a real test suite, no autocomplete. It passes or it fails. That score means the same thing next month.
-
02 — Incident
Investigate a host under a clock
A simulated box in the browser. We record which evidence they opened — not how clever the write-up sounded.
-
03 — Report
See who is drifting, by name
Two people trending down is a conversation this week. Not a vibe you notice after Friday’s page.
The product
What they open. What you read.
Incident bay
A simulated host. The clock is running. Nothing touches your infrastructure.
Kaerus
API container restarting every few minutes
mediumElapsed
04:18
ssh deploy@web-01
simulated host
Connected to web-01. Type help.
$ docker ps
api Restarting (1) 3 seconds ago
$ dmesg | tail
oom-kill: container=api, constraint=memory. Exit 137.
$ cat /etc/docker/compose.yml
mem_limit: 128m
deploy@web-01:~$
The cgroup limit is 128m. The kernel is killing the process…
Submit diagnosisDrill
The failing check is already on the page. The assistant is not.
Binary search never returns
mediumPrompt
binary_search should return the index, or -1. On some inputs it never returns — the search bounds are not always narrowing.
Checks
1 failed, 2 passed
# test_finds_present_value
.
# test_missing_value_on_narrowing_bounds
TIMEOUT — search did not return
Editor · Ruby
def binary_search(sorted_array, target)
low = 0
high = sorted_array.length - 1
while low <= high
mid = (low + high) / 2
case sorted_array[mid] <=> target
when 0 then return mid
when -1 then low = mid
when 1 then high = mid - 1
end
end
-1
end
The assistant is off. The failing line is yours to see.
Team dashboard
The manager’s view. Who is drifting, named.
Your organization
Team dashboard
Next
2 people trending down. Start an incident simulation before the next real page.
Members
12
Attempts
184
Pass rate
76%
Trending down
2
Company performance
Skill trend
-
Elena Voss
14 graded attempts
82% → 61%
Trending down -
Jonah Park
11 graded attempts
74% → 58%
Trending down -
Amira Cole
16 graded attempts
71% → 79%
+8 -
Sam Okonkwo
19 graded attempts
88% → 90%
+2
Why it is off
With the assistant open, you measure the model.
With it closed, you measure the engineer. An outage will take the second measurement either way. Kaerus takes it first, on purpose, on a clock, with a record of what they actually opened.
Unaided skill
You are here
Next incident
Who this is for
If this is already happening on your team
1:1s
You stopped asking if they can still debug
The question feels rude. The report is whether they observed the OOM line before they restarted the box. That conversation doesn’t need a guess.
After the page
The postmortem already has the ending
You already know who froze and who didn’t. A weekly sim is cheaper than another Friday you can narrate in advance. The dashboard flags drift before the next deploy does.
Get started
Open a team today. No card.
Drills, incidents, and the dashboard are on. We’ll talk about a price after you’ve used it — not before.
Know before an outage tells you