1
00:00:00,099 --> 00:00:05,059
You know, there is this absolute silent
killer when you build multi agent setups

2
00:00:05,099 --> 00:00:09,880
where a nested subagent gets stuck in a
loop, burns through twenty thousand tokens

3
00:00:09,920 --> 00:00:14,880
in three seconds, and you do not even
realize it until your root session crashes.

4
00:00:17,097 --> 00:00:18,617
Oh, I have been there!

5
00:00:18,897 --> 00:00:22,958
You check the terminal and it is just
money evaporating into the void.

6
00:00:23,617 --> 00:00:28,257
Big shoutout to Jellypod by the way, for
helping us make this daily show a reality

7
00:00:28,317 --> 00:00:32,317
every morning so we can cover fixes for
exactly that kind of nightmare.

8
00:00:32,886 --> 00:00:33,546
Absolutely.

9
00:00:34,046 --> 00:00:36,206
Huge thanks to Jellypod.

10
00:00:36,246 --> 00:00:42,226
And speaking of agent nightmare fixes,
OpenAI dropped Codex zero point one five one

11
00:00:42,306 --> 00:00:45,667
point zero today, and it hits that exact
problem head on.

12
00:00:46,007 --> 00:00:48,067
Okay, so pull back the curtain.

13
00:00:48,807 --> 00:00:53,827
What actually changed under the hood with
subagent budgets in pull request four one

14
00:00:53,907 --> 00:00:54,588
one eight three?

15
00:00:55,141 --> 00:01:00,482
So previously, if your root prompt spawned
a subagent to go inspect a directory or

16
00:01:00,542 --> 00:01:06,701
refactor a module, that subagent
maintained its own isolated local token counter.

17
00:01:06,781 --> 00:01:08,981
It had its own little sandbox world.

18
00:01:09,681 --> 00:01:15,321
But now, in zero point one five one point
zero, every single token consumed by a

19
00:01:15,382 --> 00:01:20,121
subagent is aggregated directly against
the parent session root goal budget.

20
00:01:20,507 --> 00:01:25,426
Wait, so if I give my main task a fifty
thousand token budget,

21
00:01:25,466 --> 00:01:29,026
and my main agent spins up three helper
subagents...

22
00:01:29,333 --> 00:01:33,613
They are all drawing from that exact same
fifty thousand token pool.

23
00:01:33,703 --> 00:01:38,373
The moment the sum of root tokens plus
subagent tokens hits your ceiling,

24
00:01:38,485 --> 00:01:41,013
boom, execution stops.

25
00:01:41,629 --> 00:01:46,089
Mm, okay, that is a massive safety net
against runaway loops,

26
00:01:46,670 --> 00:01:50,569
but there is a real catch there for
developers, right?

27
00:01:51,090 --> 00:01:55,829
If you set a tight budget, say ten
thousand tokens, and you ask for a complex

28
00:01:55,949 --> 00:02:01,029
codebase refactor, your subagent might be
halfway through writing a critical class

29
00:02:01,369 --> 00:02:05,409
and get hard killed right mid sentence
because the root budget ran out.

30
00:02:06,030 --> 00:02:06,569
Exactly.

31
00:02:07,009 --> 00:02:09,730
You have to recalibrate your root goal
limits now.

32
00:02:10,309 --> 00:02:15,210
You cannot just copy over your old single
agent numbers because subagents are no

33
00:02:15,329 --> 00:02:18,109
longer getting a free pass on token
consumption.

34
00:02:18,507 --> 00:02:19,806
That makes total sense.

35
00:02:20,446 --> 00:02:24,026
Now what about the other huge headline
feature in this release,

36
00:02:24,266 --> 00:02:28,966
the extension middleware for Model Context
Protocol, or MCP tools?

37
00:02:29,369 --> 00:02:34,690
Oh, pull request four one two zero two is
huge if you run tools that vomit massive

38
00:02:34,750 --> 00:02:35,250
payloads.

39
00:02:35,969 --> 00:02:41,750
Extensions registered in Codex zero point
one five one point zero can now inspect,

40
00:02:42,250 --> 00:02:48,609
sanitize, or completely replace raw JSON
and text outputs from MCP tools

41
00:02:49,029 --> 00:02:52,250
before those payloads ever reach the model
context window.

42
00:02:53,317 --> 00:02:53,617
Wait!

43
00:02:53,977 --> 00:02:57,317
Client side interception before it hits
the context?

44
00:02:57,708 --> 00:02:58,588
Yes!

45
00:02:58,641 --> 00:03:04,428
Think about what happens when an MCP
database tool returns a fifty kilobyte JSON

46
00:03:04,508 --> 00:03:09,468
response with a thousand rows, but your
prompt only needed three fields from two

47
00:03:09,528 --> 00:03:09,948
records.

48
00:03:10,050 --> 00:03:15,068
Previously, that entire fifty kilobyte
blob was dumped straight into your context

49
00:03:15,114 --> 00:03:18,428
window, burning context and costing real
money.

50
00:03:19,442 --> 00:03:25,942
Or worse, returning an internal API key or
bearer token inside a debug header that

51
00:03:25,982 --> 00:03:28,003
gets fed straight into the model prompt.

52
00:03:28,471 --> 00:03:29,091
Exactly.

53
00:03:29,591 --> 00:03:34,411
Now, with middleware extensions, you can
write a tiny filter script that intercepts

54
00:03:34,451 --> 00:03:39,592
the tool output, strips out sensitive
headers, prunes away seventy percent of the

55
00:03:39,731 --> 00:03:45,211
useless JSON fields, and forwards only the
lean, clean payload to the model.

56
00:03:45,776 --> 00:03:50,116
That is incredible for both security and
prompt efficiency.

57
00:03:50,636 --> 00:03:53,576
I love that control sitting right on the
client side.

58
00:03:54,197 --> 00:04:00,397
Beyond middleware and token budgets, I saw
a whole cluster of security and sandbox

59
00:04:00,476 --> 00:04:02,177
fixes in the changelog.

60
00:04:02,236 --> 00:04:06,237
What was going on with the TUI session
paths and permission profiles?

61
00:04:06,661 --> 00:04:13,141
Okay, so pull requests four one one nine
two and four one two zero nine fixed a

62
00:04:13,181 --> 00:04:14,981
couple of subtle sandbox leaks.

63
00:04:15,681 --> 00:04:21,401
In older versions, if you changed
directories using slash cd inside an interactive

64
00:04:21,462 --> 00:04:27,021
terminal session, it could actually weaken
your preconfigured sandbox boundaries or

65
00:04:27,062 --> 00:04:30,001
lose track of restored permission profiles
across turns.

66
00:04:31,361 --> 00:04:31,902
Yikes.

67
00:04:32,301 --> 00:04:37,861
So you change directory to inspect a
subfolder and suddenly your permission profile

68
00:04:38,042 --> 00:04:40,061
loses its strict boundary enforcement?

69
00:04:40,375 --> 00:04:40,935
Right.

70
00:04:41,015 --> 00:04:46,135
In zero point one five one point zero,
permission profiles are preserved perfectly

71
00:04:46,192 --> 00:04:51,895
across TUI turns, and path resolution now
respects the remote executor actual home

72
00:04:51,951 --> 00:04:57,895
directory, operating system, and path
semantics under pull requests four one two

73
00:04:57,975 --> 00:05:01,495
zero four and four one two zero seven.

74
00:05:01,548 --> 00:05:07,015
So Windows backslashes versus Linux
forward slashes won't bypass your deny read

75
00:05:07,068 --> 00:05:07,895
rules anymore.

76
00:05:08,174 --> 00:05:10,714
And what about the Guardian security
layer?

77
00:05:11,174 --> 00:05:14,815
I saw a bug fix there around stale
classification cache.

78
00:05:15,327 --> 00:05:18,327
That was pull request four one one nine
six.

79
00:05:18,967 --> 00:05:23,687
If you modified your security approvals or
updated permission states mid session,

80
00:05:24,247 --> 00:05:29,048
the Guardian classification system was
sometimes holding onto cached approvals from

81
00:05:29,107 --> 00:05:34,548
before the change, meaning an action you
just revoked could still execute because of

82
00:05:34,587 --> 00:05:36,107
a stale cached decision.

83
00:05:36,867 --> 00:05:41,388
Now, any permission state change revokes
those cached classifications immediately.

84
00:05:41,777 --> 00:05:42,517
Good catch.

85
00:05:43,237 --> 00:05:45,737
Now, what about developer quality of life?

86
00:05:46,157 --> 00:05:49,097
Anything for people whose tools take
forever to boot up?

87
00:05:50,136 --> 00:05:50,956
Oh, absolutely.

88
00:05:51,496 --> 00:05:57,456
Pull request four one one nine nine added
a configurable startup grace period for

89
00:05:57,536 --> 00:05:59,896
discovering optional MCP servers.

90
00:06:00,596 --> 00:06:04,196
If you have an MCP server that takes five
seconds to initialize,

91
00:06:04,676 --> 00:06:08,496
Codex won't just fail or hang your entire
command line interface startup.

92
00:06:09,097 --> 00:06:11,536
You can set the grace period to wait
gracefully.

93
00:06:11,863 --> 00:06:18,142
And pull request four one two zero eight
fixes plugin catalog aggregation!

94
00:06:18,202 --> 00:06:21,643
So if you have a broken project
marketplace in one repository,

95
00:06:22,022 --> 00:06:25,642
it won't crash or suppress your valid
local plugins anymore.

96
00:06:26,003 --> 00:06:29,622
It just flags the invalid one and loads
everything else cleanly.

97
00:06:30,182 --> 00:06:30,882
Precisely.

98
00:06:31,382 --> 00:06:36,182
So if you are upgrading to Codex zero
point one five one point zero today,

99
00:06:36,262 --> 00:06:39,642
review your root goal token budgets for
multi agent workflows,

100
00:06:40,262 --> 00:06:44,722
write a custom extension filter if your
MCP tools return bloated JSON,

101
00:06:45,302 --> 00:06:47,622
and enjoy the tighter sandbox enforcement.

102
00:06:48,146 --> 00:06:49,506
Solid release overall.

103
00:06:49,946 --> 00:06:51,407
Good stuff from the OpenAI team.

104
00:06:51,906 --> 00:06:52,647
Talk to you all tomorrow!