1
00:00:00,000 --> 00:00:03,108
Welcome to the show - I'm Lachlan Reed
with James Turner,

2
00:00:03,108 --> 00:00:06,610
and James, I love this because it sounds
like autocomplete,

3
00:00:06,610 --> 00:00:07,807
but it's really not.

4
00:00:07,807 --> 00:00:12,359
You type /ultrareview, and instead of
some little in-editor nudge,

5
00:00:12,359 --> 00:00:16,186
it kicks off multiple reviewer agents in
parallel up in the cloud.

6
00:00:16,186 --> 00:00:17,384
Then you just...

7
00:00:17,384 --> 00:00:18,100
keep coding.

8
00:00:18,100 --> 00:00:19,695
Open other commands.

9
00:00:19,695 --> 00:00:20,650
Go make a coffee.

10
00:00:20,650 --> 00:00:24,508
Apparently you can even close the
terminal while it's still chewing through

11
00:00:24,508 --> 00:00:28,311
your branch, which feels a bit like
leaving a team of very judgmental

12
00:00:28,311 --> 00:00:30,467
interns alone in your workshop.

13
00:00:30,467 --> 00:00:35,684
The "close the
terminal" part is the token that changes

14
00:00:35,684 --> 00:00:36,549
it for me.

15
00:00:36,549 --> 00:00:39,220
If I can shut the terminal and the review
keeps running,

16
00:00:39,220 --> 00:00:43,647
that is NOT the same category as a
chatbot hovering over my shoulder.

17
00:00:43,647 --> 00:00:44,963
That's a job queue.

18
00:00:44,963 --> 00:00:49,236
That's basically saying, "we sent your
code to a cloud review pipeline;

19
00:00:49,236 --> 00:00:51,556
come back in five to ten minutes."Exactly.

20
00:00:51,556 --> 00:00:53,306
And the entry points are dead simple.

21
00:00:53,306 --> 00:00:57,621
/ultrareview with no arguments reviews
your current git branch.

22
00:00:57,621 --> 00:00:58,779
Easy as.

23
00:00:58,779 --> 00:01:03,569
Or /ultrareview followed by a PR number
targets a specific GitHub pull

24
00:01:03,569 --> 00:01:04,790
request before merge.

25
00:01:04,790 --> 00:01:09,370
So one version is, "check what I've been
hacking on locally," and the

26
00:01:09,370 --> 00:01:14,176
other is, "have a proper squiz at PR 214
before this thing lands."[responds

27
00:01:14,176 --> 00:01:16,704
quickly] Wait - PR number,
not a full link?

28
00:01:16,704 --> 00:01:22,042
That's interesting, because it makes the
pull request feel like a first-class

29
00:01:22,042 --> 00:01:24,770
object, not just text you pasted in.

30
00:01:24,770 --> 00:01:28,124
And the branch mode versus PR mode
matters too.

31
00:01:28,124 --> 00:01:32,587
A branch review is like catching dumb
mistakes before you show anyone.

32
00:01:32,587 --> 00:01:35,623
A PR review is more like, "okay,

33
00:01:35,623 --> 00:01:38,480
this is entering the blast radius."Yeah,

34
00:01:38,480 --> 00:01:40,232
and the bigger idea is trust.

35
00:01:40,232 --> 00:01:42,711
If one model reads your diff once,

36
00:01:42,711 --> 00:01:44,779
you're getting one first impression.

37
00:01:44,779 --> 00:01:47,012
Maybe brilliant, maybe half asleep,

38
00:01:47,012 --> 00:01:49,246
maybe obsessed with a variable name.

39
00:01:49,246 --> 00:01:53,707
But if /ultrareview is spinning up
multiple reviewer agents in parallel,

40
00:01:53,707 --> 00:01:56,935
each coming at it from different angles -
logic,

41
00:01:56,935 --> 00:02:01,842
edge cases, security, performance - then
it's a multi-pass review.

42
00:02:01,842 --> 00:02:03,122
Different lenses.

43
00:02:03,122 --> 00:02:04,722
Different failure modes.

44
00:02:04,722 --> 00:02:09,315
That's a very different beast.
Or - and this is my pushback

45
00:02:09,315 --> 00:02:12,667
- it's just more expensive linting with
nicer packaging.

46
00:02:12,667 --> 00:02:16,569
Because "multiple agents in parallel"
sounds impressive,

47
00:02:16,569 --> 00:02:19,788
but if I still wait five to ten minutes
and get back a polished list

48
00:02:19,788 --> 00:02:22,862
of things my linter or test suite
should've caught,

49
00:02:22,862 --> 00:02:28,504
then cool, you built a cloud chandelier
over a flashlight.A

50
00:02:28,504 --> 00:02:31,069
cloud chandelier is harsh...

51
00:02:31,069 --> 00:02:31,629
but fair.

52
00:02:31,629 --> 00:02:34,429
I think that's the tension here.

53
00:02:34,429 --> 00:02:37,139
Is this actually better code review,

54
00:02:37,139 --> 00:02:39,530
or is it fancy linting with a longer
queue?

55
00:02:39,530 --> 00:02:43,116
Because if it's just style nags in a
trench coat,

56
00:02:43,116 --> 00:02:43,676
I'm out.

57
00:02:43,676 --> 00:02:47,346
But if it genuinely finds the weird
branch interaction,

58
00:02:47,346 --> 00:02:51,909
the missed auth check, the performance
footgun - stuff a single pass

59
00:02:51,909 --> 00:02:56,338
might miss - then five minutes starts
sounding pretty cheap.

60
00:02:51,909 --> 00:02:56,338
And that's

61
00:02:56,338 --> 00:03:00,940
the thing I'm listening for: not "did it
say something," but "did multiple

62
00:03:00,940 --> 00:03:05,460
parallel reviewers produce something I'd
trust more than one model's vibe?"

63
00:03:05,460 --> 00:03:10,160
Because those are very different
promises.

64
00:03:05,460 --> 00:03:10,160
So here's the mechanism that

65
00:03:10,160 --> 00:03:14,937
matters most: findings are independently
reproduced and verified before

66
00:03:14,937 --> 00:03:15,621
they appear.

67
00:03:15,621 --> 00:03:18,419
That wording is doing a lot of work.

68
00:03:18,419 --> 00:03:21,689
The system is trying to surface confirmed
issues,

69
00:03:21,689 --> 00:03:23,280
not speculative warnings.

70
00:03:23,280 --> 00:03:27,512
Not "hey,
maybe there's a bug?" but "we checked,

71
00:03:27,512 --> 00:03:31,426
and this actually
reproduces."The word

72
00:03:27,512 --> 00:03:31,426
reproduced

73
00:03:31,426 --> 00:03:32,906
is HUGE there.

74
00:03:32,906 --> 00:03:36,372
Reproduced means it's not just one model
having a funny feeling.

75
00:03:36,372 --> 00:03:38,882
It's the difference between your mate
saying,

76
00:03:38,882 --> 00:03:42,744
"that bike sounds crook," and another
mechanic actually riding it,

77
00:03:42,744 --> 00:03:45,215
hearing the same knock,
and pointing to the loose bit.

78
00:03:45,215 --> 00:03:48,360
Different level of confidence entirely.

79
00:03:45,215 --> 00:03:48,360
Right.

80
00:03:48,360 --> 00:03:50,070
In a normal single-agent flow,

81
00:03:50,070 --> 00:03:51,984
one model can just go off the rails.

82
00:03:51,984 --> 00:03:53,342
It can miss a real bug.

83
00:03:53,342 --> 00:03:54,697
It can fixate on style.

84
00:03:54,697 --> 00:03:57,802
It can hallucinate a problem because a
pattern looked suspicious.

85
00:03:57,802 --> 00:04:02,523
/ultrareview is trying to reduce that by
having agents specialize and

86
00:04:02,523 --> 00:04:03,303
then cross-check.

87
00:04:03,303 --> 00:04:06,733
So instead of one pass that's broad and a
little flaky,

88
00:04:06,733 --> 00:04:11,185
you've got multiple passes with some
internal skepticism built in.

89
00:04:06,733 --> 00:04:11,185
Let

90
00:04:11,185 --> 00:04:12,434
me try to explain it back.

91
00:04:12,434 --> 00:04:19,863
A standard one-shot review is
like asking one very smart dev

92
00:04:19,863 --> 00:04:21,857
to skim the diff after lunch.

93
00:04:21,857 --> 00:04:24,811
Maybe they spot the nasty edge case,

94
00:04:24,811 --> 00:04:29,279
maybe they don't, maybe they spend ten
minutes on tabs versus spaces.

95
00:04:29,279 --> 00:04:34,876
Ultrareview is more like sending the same
diff to a little committee

96
00:04:34,876 --> 00:04:39,269
- one looking for logic gaps,
another for security weirdness,

97
00:04:39,269 --> 00:04:45,044
another for performance traps - and then
only surfacing stuff that survives

98
00:04:45,044 --> 00:04:46,064
the committee scrap.

99
00:04:46,064 --> 00:04:49,632
Is that close?

100
00:04:46,064 --> 00:04:49,632
Pretty close.

101
00:04:49,632 --> 00:04:54,908
The key add-on is that the
five-to-ten-minute runtime makes the cloud

102
00:04:54,908 --> 00:04:56,028
part real.

103
00:04:56,028 --> 00:04:59,058
This is deliberately a background task,

104
00:04:59,058 --> 00:05:01,931
not conversational back-and-forth inside
the editor.

105
00:05:01,931 --> 00:05:03,770
You don't sit there nudging it like,

106
00:05:03,770 --> 00:05:07,123
"what about line 48?" You launch it,

107
00:05:07,123 --> 00:05:09,120
it goes off and does heavy work,

108
00:05:09,120 --> 00:05:12,748
and it returns later.

109
00:05:09,120 --> 00:05:12,748
And I actually like that separation.

110
00:05:12,748 --> 00:05:16,020
When a tool replies
instantly,

111
00:05:16,020 --> 00:05:18,410
I start treating it like a mate on chat.

112
00:05:18,410 --> 00:05:20,646
Handy, but a bit casual.

113
00:05:20,646 --> 00:05:23,434
A five-minute background run says,

114
00:05:23,434 --> 00:05:29,399
"nah, this is a process." It asks you to
use it on the bits that matter.

115
00:05:23,434 --> 00:05:29,399
See,

116
00:05:29,399 --> 00:05:30,678
I half-agree.

117
00:05:30,678 --> 00:05:34,186
"Verified findings" definitely changes
confidence.

118
00:05:34,186 --> 00:05:37,859
But it can also hide uncertainty behind a
cleaner interface.

119
00:05:37,859 --> 00:05:41,684
Because all verification systems still
have thresholds,

120
00:05:41,684 --> 00:05:43,762
judgments, edge cases.

121
00:05:43,762 --> 00:05:48,223
Just because the output looks polished
doesn't mean the messy uncertainty

122
00:05:48,223 --> 00:05:48,871
vanished.

123
00:05:48,871 --> 00:05:52,878
It might just be better concealed.

124
00:05:48,871 --> 00:05:52,878
That's a good shove.

125
00:05:52,878 --> 00:05:56,762
You're saying the phrase confirmed issues
can make people lower their

126
00:05:56,762 --> 00:05:59,034
guard too much.

127
00:05:56,762 --> 00:05:59,034
Exactly.

128
00:05:59,034 --> 00:06:01,347
"Confirmed" is powerful language.

129
00:06:01,347 --> 00:06:05,507
I'm not saying it's fake - I'm saying
developers might hear confirmed

130
00:06:05,507 --> 00:06:07,484
and mentally skip review.

131
00:06:07,484 --> 00:06:09,321
And that would be a mistake.

132
00:06:09,321 --> 00:06:10,953
Higher confidence?

133
00:06:10,953 --> 00:06:11,313
Sure.

134
00:06:11,313 --> 00:06:12,911
Final authority?

135
00:06:12,911 --> 00:06:15,560
No chance.

136
00:06:12,911 --> 00:06:15,560
Yeah.

137
00:06:15,560 --> 00:06:18,915
So maybe the real win isn't certainty.

138
00:06:18,915 --> 00:06:21,383
Maybe it's better triage.

139
00:06:21,383 --> 00:06:23,859
If the thing comes back with fewer,

140
00:06:23,859 --> 00:06:28,562
more solid findings, that saves me from
drowning in speculative nonsense.

141
00:06:28,562 --> 00:06:33,826
That alone feels different from a single
AI pass spraying warnings everywhere.