ancha.

Watch an AI agent get tested.

One real run, step by step.

1

PagerDuty
pages it
Datadog
finds the cause
AI on-call agent Claude Haiku 5.5
ships the fix
GitHub
updates the team
Slack
15tools 232lines of code < 1 hwritten by Claude
Same agent not one line changed
Productionoff limits
SlackGitHubDatadogPagerDuty
acme · simulatedsafe to break
SlackGitHubDatadogPagerDuty
runbook code history monitor on-call rota
13:55
Bad change ships 32 → 3
13:56
Orders fail
14:00
Alert fires
14:00
Agent paged
14:00
Slack lights up
No script, no button: the world's rules did this
agent world
0:001:002:003:004:00
0:06 – 0:14 took the page · found the cause · merged a fix
0:29 fix deployed
1:30 · Ben B"Status? Reply in this thread."
2:17 – 4:21 watches the monitor, not Slack
watches the monitor, not Slack
2:30 recovered · page closed
4:23 "recovered", in the channel
Version 1 · 3 copies 0 / 3 passed
agent reads Slack
0:06never again
Ben asks
B1:30
agent posts
0:124:23
0:001:002:003:004:00
✗ answered_request_thread expected ≥1 reply from the agent in Ben's thread actual 2 messages in the channel, 0 in the thread
#orders-incident
B
Ben 14:01:30

Status request: what is the current customer impact, what are you doing, and when is the next update? Reply in this thread.

0 replies
A second AI, grading the reply on its own: fail
+before every step→Slack
Version 2 · 3 copies 0 / 3 passed
#orders-incident
B
Ben 14:01:30

Status request: what is the current customer impact, what are you doing, and when is the next update? Reply in this thread.

1 reply
incident-bot 14:02:16

Status: Not recovered yet.
Impact: Orders requests are still failing; the monitor is still in Alert.
Action: the fix is merged and deployed. Watching the backlog drain. Recovery not confirmed.
Next update: in about 1 minute.

Bad deployroll back, confirm recovery
False alarmleave production alone
Status requestfix it, answer Ben
Disk fullhand it to the DB person no reassign tool
Noisy alarmmute it for a few minutes no mute tool
Coding agent alone
With Ancha
Accounts
no real accounts; can't practice on production
copies of all four, one command
The incident
must fake four tools that react to each other
the world runs on its own rules
People and time
someone has to ask at 1:30, on time
teammates show up on schedule
The grade
writes the test, grades itself
outside referee, replay of every step
the agent271
Ancha's world6,571 · 24×
test code we wrote0
Our Oct 28 attempt by hand: test code 3× the agent, with 7 bugs that each reported a false result.