1
00:00:08,177 --> 00:00:12,318
You know, last week we were digging deep
into the terminal backend fixes in Grok

2
00:00:12,418 --> 00:00:18,757
Build 1.0.13, looking at self healing
retries and background task monitors.

3
00:00:18,817 --> 00:00:23,758
But this week, xAI completely shifted
gears toward the user facing experience,

4
00:00:24,277 --> 00:00:28,798
specifically around Grok Bot and how these
AI teammates interact with the real

5
00:00:28,917 --> 00:00:29,277
world.

6
00:00:29,473 --> 00:00:33,333
And before we jump into the mechanics, a
quick shoutout, today's episode is brought

7
00:00:33,392 --> 00:00:35,132
to you by Jellypod AI.

8
00:00:36,052 --> 00:00:39,812
I actually spent the last few days testing
out the brand new native X connector in

9
00:00:39,892 --> 00:00:42,592
Grok Bot for my own real time tech news
monitoring.

10
00:00:43,412 --> 00:00:48,212
And I, I gotta say, the difference between
native feed access and standard web

11
00:00:48,272 --> 00:00:50,253
search scraping is night and day.

12
00:00:50,411 --> 00:00:55,151
See, previously, if you wanted an agent to
track brand mentions or summarize

13
00:00:55,272 --> 00:01:00,712
breaking live threads on X, it was either
relying on search index latency or you had

14
00:01:00,712 --> 00:01:03,591
to go through this tedious third party API
setup.

15
00:01:04,411 --> 00:01:05,212
Now?

16
00:01:05,292 --> 00:01:08,472
You just open Grok Bot, click sign in with
the X connector,

17
00:01:08,871 --> 00:01:13,451
and xAI automatically provisions a
developer account for you if you do not already

18
00:01:13,511 --> 00:01:13,892
have one.

19
00:01:14,472 --> 00:01:19,052
Plus, paid Grok Bot users get free X API
credits right out of the box.

20
00:01:19,162 --> 00:01:23,443
So, instead of a bot pretending to browse
a webpage, it is querying posts,

21
00:01:23,882 --> 00:01:28,742
reading your timeline, checking mentions,
and pulling live conversations straight

22
00:01:28,822 --> 00:01:30,122
through dedicated tools.

23
00:01:30,682 --> 00:01:34,142
When I had my bot tracking developer
sentiment around recent model releases,

24
00:01:34,583 --> 00:01:38,622
it was pulling structured mention graphs
and live thread context instantly,

25
00:01:39,022 --> 00:01:41,822
without getting rate limited or stuck on
missing elements.

26
00:01:42,054 --> 00:01:45,554
But here is where the desktop app layer
gets really interesting.

27
00:01:46,255 --> 00:01:49,494
Because an agent can only go as far as its
authentication allows.

28
00:01:50,114 --> 00:01:54,234
That brings us to the updated Agent
Computer takeover interface.

29
00:01:54,295 --> 00:01:58,815
Remember, every Grok Bot runs on its own
persistent cloud VM with a browser,

30
00:01:59,315 --> 00:02:00,555
filesystem, and terminal.

31
00:02:00,742 --> 00:02:03,162
So what happens when your bot hits a
strict wall?

32
00:02:03,722 --> 00:02:07,182
Like a multi factor authentication prompt,
a passkey request,

33
00:02:07,742 --> 00:02:09,223
or a mandatory CAPTCHA?

34
00:02:10,083 --> 00:02:14,722
In the updated desktop app on macOS and
Windows, you open the Agent Computer view

35
00:02:14,762 --> 00:02:17,962
inside your conversation thread, click the
takeover control,

36
00:02:18,382 --> 00:02:21,322
and you are looking directly at the cloud
VM screen.

37
00:02:21,962 --> 00:02:25,742
You type in your password, pass the two
factor code or complete the CAPTCHA

38
00:02:25,802 --> 00:02:29,423
yourself, and then simply click to hand
control right back to the bot.

39
00:02:29,557 --> 00:02:34,017
It completely eliminates that clunky back
and forth where an agent fails silently or

40
00:02:34,458 --> 00:02:36,437
dumps raw error logs into chat.

41
00:02:36,977 --> 00:02:40,277
The browser session stays logged in on
your shared persistent computer,

42
00:02:40,697 --> 00:02:44,677
meaning any other bot on your account can
leverage that same signed in environment

43
00:02:44,717 --> 00:02:45,638
for future tasks.

44
00:02:45,867 --> 00:02:48,887
Now, where are people actually deploying
these bots right now?

45
00:02:49,567 --> 00:02:53,287
The use cases coming out of production
environments are surprisingly operational.

46
00:02:53,968 --> 00:02:57,807
We are seeing Office Manager bots that
intake work orders across Slack,

47
00:02:58,267 --> 00:03:03,468
pull customer data from Gmail, and
navigate browser portals like ServiceTitan or Quo

48
00:03:03,527 --> 00:03:05,347
to plan capacity and book jobs.

49
00:03:05,515 --> 00:03:07,875
Or take Expense Reconciliation bots.

50
00:03:08,375 --> 00:03:12,434
They audit weekly spend, pull receipts
directly out of finance inboxes,

51
00:03:12,934 --> 00:03:17,715
match them against corporate policies,
flag exceptions, and draft follow up emails

52
00:03:17,774 --> 00:03:18,715
for missing details.

53
00:03:19,594 --> 00:03:24,114
These are not hypothetical prompt
experiments, they are actual multi system,

54
00:03:24,534 --> 00:03:28,014
multi tool workflows running in background
persistent loops.

55
00:03:28,242 --> 00:03:32,722
But running digital coworkers twenty four
seven brings up a massive operational

56
00:03:32,842 --> 00:03:35,862
bottleneck that I think a lot of teams are
bumping into right now,

57
00:03:36,262 --> 00:03:37,162
approval fatigue.

58
00:03:38,003 --> 00:03:38,543
Think about it.

59
00:03:38,802 --> 00:03:42,503
If you have three or four bots auditing
systems, drafting outreach,

60
00:03:42,822 --> 00:03:46,722
and managing tickets around the clock, you
quickly get flooded with dozens of

61
00:03:46,782 --> 00:03:48,923
notification cards asking for your green
light.

62
00:03:49,077 --> 00:03:54,237
If every minor step requires human
confirmation, you end up acting as a glorified

63
00:03:54,398 --> 00:03:58,517
router, defeat, defeating the entire
purpose of an autonomous teammate.

64
00:03:59,157 --> 00:04:03,557
But if you auto approve everything
blindly, a bot might execute an unintended

65
00:04:03,577 --> 00:04:07,217
production change or send unvetted
external communications.

66
00:04:07,997 --> 00:04:12,017
That is why precise hook design and clear
permission boundaries are so vital.

67
00:04:12,342 --> 00:04:14,823
So how do you actually structure a safe
handoff?

68
00:04:15,502 --> 00:04:19,142
The golden rule here is to always start
with read and prepare tasks.

69
00:04:19,762 --> 00:04:23,643
Have your bot research prospect lists,
score leads against an ideal customer

70
00:04:23,742 --> 00:04:29,162
profile, or draft outreach, but leave
every final action as an unexecuted review

71
00:04:29,202 --> 00:04:30,263
list in your inbox.

72
00:04:30,532 --> 00:04:34,693
Once you observe the bot executing that
exact sequence reliably across multiple

73
00:04:34,753 --> 00:04:38,952
runs, you can save that multi step path as
a persistent skill or routine.

74
00:04:39,452 --> 00:04:43,753
But keep this rule in mind, create a
routine only when retries and failure cases are

75
00:04:43,772 --> 00:04:44,213
defined.

76
00:04:44,469 --> 00:04:45,870
That distinction is key.

77
00:04:46,549 --> 00:04:51,429
If a bot does not know what to do when a
web portal times out or a API endpoint

78
00:04:51,489 --> 00:04:56,149
returns a rate limit, it should never be
running on an automated schedule with live

79
00:04:56,229 --> 00:04:57,929
execution privileges.

80
00:04:57,989 --> 00:05:02,430
Define the failure behavior first, lock
sensitive external actions behind approval

81
00:05:02,529 --> 00:05:05,590
gates, and let the bot handle the heavy
lifting in the background.

82
00:05:05,844 --> 00:05:09,744
Alright, that is the latest on Grok Bot
and native X workflows.

83
00:05:09,784 --> 00:05:12,924
Give those takeover controls a spin, and I
will talk to you all next week.