It says done.
Is it? 01
Who reviews
the reviewer? 02
Fast to build.
Slow to trust? 03
What did
the retry cost? 04
What changed
in production? 05
More context.
Less signal? 06
Would you ship
without checking? 07
Can it fail
quietly? 08
Who owns
the mistake? 09
Where did
the state go? 10
Did the tool work,
or just return OK? 11
What is your
stop condition? 12
Does your eval
look like your users? 13
A demo is
one good run. 14
What survives
a model update? 15
How long until
you notice? 16
Did we automate
the wrong step? 17
Who cleans up
after the agent? 18
When should
it ask a human? 19
What would
change your mind? 20
Can you explain
the diff? 21
Is the test
testing anything? 22
What did
it forget? 23
One agent.
How many permissions? 24
Why another
agent? 25
Can you replay
the failure? 26
What happens
on run 100? 27
Does it need
write access? 28
Who has time
to read the output? 29
Are you faster,
or just busier? 30
Where is
the evidence? 31
What did
we actually save? 32
Is this a fix
or another layer? 33
What should
stay manual? 34
Can it undo
what it did? 35
How stale
is that memory? 36
Would a script
be enough? 37
Who decides
what good means? 38
What breaks
when nobody watches? 39
Can we debug
the handoff? 40
How many retries
is a loop? 41
Useful output
or more output? 42
Whose data
is in the context? 43
What deserves
your attention? 44
Fewer prompts.
Better boundaries? 45
Does it know
when it is stuck? 46
What is the cost
per useful result? 47
Bring the failure.
28 October. 48
