1
00:00:08,117 --> 00:00:12,697
So, I have been thinking a lot about how
we actually build software with AI agents

2
00:00:12,757 --> 00:00:13,117
right now.

3
00:00:13,777 --> 00:00:19,018
Last time, we talked about Grok 4.6
reasoning levels and how xAI opened up broader

4
00:00:19,057 --> 00:00:19,737
plan tiers.

5
00:00:20,537 --> 00:00:24,558
But, er, something shifted this past week
that goes way beyond raw benchmarks,

6
00:00:25,138 --> 00:00:28,738
and honestly, it completely changes how I
think about delegating actual work.

7
00:00:28,750 --> 00:00:33,150
For years, the standard way to teach an AI
model a complex workflow was,

8
00:00:33,198 --> 00:00:35,630
well, writing massive prompt templates.

9
00:00:35,690 --> 00:00:38,910
You would sit down and spend hours
crafting system prompts,

10
00:00:38,944 --> 00:00:43,550
trying to explain every single edge case,
or you had to build custom Model Context

11
00:00:43,603 --> 00:00:47,950
Protocol servers and developer APIs just
so the agent could click a single button on

12
00:00:48,030 --> 00:00:48,670
a screen.

13
00:00:48,830 --> 00:00:53,550
But xAI just introduced what they are
calling Routines in Grok Bot,

14
00:00:53,570 --> 00:00:55,790
and it flips that entire paradigm on its
head.

15
00:00:55,871 --> 00:00:59,851
Instead of writing out instructions, you
just do the job once while the bot watches.

16
00:01:00,311 --> 00:01:01,051
That is it.

17
00:01:01,471 --> 00:01:04,351
You run through your manual web workflow,
you log into the portal,

18
00:01:04,672 --> 00:01:09,111
click the buttons, download the report,
and the bot records those UI actions and

19
00:01:09,151 --> 00:01:12,292
turns them directly into a repeatable,
background job.

20
00:01:12,751 --> 00:01:14,472
No custom API required.

21
00:01:14,531 --> 00:01:16,011
No MCP server setup.

22
00:01:16,292 --> 00:01:20,371
It is literally monkey see, monkey do for
autonomous cloud agents.

23
00:01:20,596 --> 00:01:22,576
Now, how does that actually work under the
hood?

24
00:01:23,336 --> 00:01:27,636
According to xAI documentation, every Grok
Bot runs inside its own dedicated,

25
00:01:27,897 --> 00:01:29,476
always on cloud computer.

26
00:01:30,217 --> 00:01:33,817
This is not just a stateless API runner
that spins up for three seconds and

27
00:01:33,896 --> 00:01:34,456
vanishes.

28
00:01:35,016 --> 00:01:39,236
It is a full virtual environment with
native browser and terminal access.

29
00:01:39,996 --> 00:01:43,856
And because it runs persistently in the
cloud, your messaging threads sync

30
00:01:43,936 --> 00:01:46,956
seamlessly across desktop and mobile iOS
devices.

31
00:01:47,696 --> 00:01:50,876
You can kick off a task on your Mac at the
office, walk away,

32
00:01:51,257 --> 00:01:54,976
and pick up the exact same live thread on
your phone while grabbing lunch.

33
00:01:55,042 --> 00:01:59,842
And, er, this is where the contrast with
raw language model upgrades gets really

34
00:01:59,889 --> 00:02:00,482
interesting.

35
00:02:00,626 --> 00:02:04,322
High reasoning benchmarks are great for
isolated code generation,

36
00:02:04,386 --> 00:02:04,802
sure.

37
00:02:04,882 --> 00:02:10,082
But the actual bottleneck in enterprise
automation has always been legacy systems.

38
00:02:10,226 --> 00:02:15,362
Most internal web tools inside large
companies do not have neat GraphQL endpoints or

39
00:02:15,419 --> 00:02:16,882
public REST APIs.

40
00:02:16,962 --> 00:02:20,402
They are old, clunky, and locked behind
web forms.

41
00:02:20,896 --> 00:02:26,056
With non API browser navigation, Grok Bot
does not care that your internal vendor

42
00:02:26,176 --> 00:02:27,656
portal was built in 2008.

43
00:02:28,276 --> 00:02:33,097
It can open the browser, navigate the raw
DOM, log in, fill out forms,

44
00:02:33,436 --> 00:02:36,036
and pull data just like a human worker
sitting at a desk.

45
00:02:36,083 --> 00:02:38,563
Take a real world example happening right
now in production.

46
00:02:38,750 --> 00:02:43,155
An operations team set up a single Grok
Bot routine to monitor three separate legacy

47
00:02:43,155 --> 00:02:44,883
logistics websites every morning.

48
00:02:45,063 --> 00:02:49,763
The bot logs into all three portals, cross
references invoice tables across the

49
00:02:49,830 --> 00:02:54,483
sites, checks for discrepancies, and then
drafts flag alerts directly into the team

50
00:02:54,536 --> 00:02:55,283
Slack channel.

51
00:02:55,459 --> 00:02:59,443
Zero custom API integrations were written
by the engineering team.

52
00:02:59,555 --> 00:03:04,403
That entire pipeline was set up in minutes
by just demonstrating the task once.

53
00:03:04,557 --> 00:03:07,296
Think about what that unlocks for small
engineering teams.

54
00:03:07,856 --> 00:03:12,136
You suddenly have digital teammates
capable of handling the messy operational glue

55
00:03:12,196 --> 00:03:16,616
work that used to devour hours of manual
developer time every single week.

56
00:03:16,746 --> 00:03:20,946
But, uh, whenever we talk about autonomous
agents logging into services and running

57
00:03:20,986 --> 00:03:25,106
tools on their own, the immediate wall
everyone hits is security.

58
00:03:25,186 --> 00:03:29,426
And let us be real, the habit most
developers developed over the last year was...

59
00:03:30,167 --> 00:03:30,646
terrible.

60
00:03:31,206 --> 00:03:36,087
People were pasting raw API keys, session
tokens, and database credentials directly

61
00:03:36,106 --> 00:03:40,167
into system prompts or temporary
environment variables just to get a demo working.

62
00:03:40,167 --> 00:03:44,727
That is why the new Secrets management
system in Grok Build is such a huge

63
00:03:44,785 --> 00:03:45,767
structural deal.

64
00:03:45,954 --> 00:03:51,127
It completely isolates third party API
credentials outside the agent prompt and code

65
00:03:51,194 --> 00:03:52,247
space entirely.

66
00:03:52,427 --> 00:03:56,887
The model can reference a credential name
to perform an authenticated task,

67
00:03:56,907 --> 00:04:01,527
but it never actually sees or leaks the
underlying secret string in its raw output

68
00:04:01,580 --> 00:04:02,887
or transcript logs.

69
00:04:03,100 --> 00:04:07,927
It solves one of the biggest security
nightmares in agentic automation with a clean,

70
00:04:07,983 --> 00:04:09,447
dedicated credential vault.

71
00:04:09,537 --> 00:04:14,378
Combine those secure secrets with xAI's
new Business Connectors and grok dot me

72
00:04:14,438 --> 00:04:18,517
publishing, and you can build full scale
operational applications on the fly.

73
00:04:19,218 --> 00:04:22,397
You can tell Grok Build to pull live
business metrics through a connector,

74
00:04:22,897 --> 00:04:27,697
format the output into a live, filterable
web dashboard, attach a custom domain,

75
00:04:28,058 --> 00:04:28,997
and host it live.

76
00:04:29,617 --> 00:04:32,317
And if you need to take that code into a
full developer workflow,

77
00:04:32,837 --> 00:04:36,857
you can export the whole project straight
to a GitHub repository with one command.

78
00:04:37,117 --> 00:04:40,757
Now, if you are looking at this from an
enterprise compliance perspective,

79
00:04:41,217 --> 00:04:44,717
you might wonder how this scales beyond
individual developer machines.

80
00:04:45,437 --> 00:04:51,017
Well, xAI release notes show that Grok 4.6
has expanded rapidly across all major

81
00:04:51,117 --> 00:04:52,297
cloud providers this month.

82
00:04:52,797 --> 00:04:56,718
It is now live on Amazon Bedrock, Google
Enterprise Agent Platform,

83
00:04:57,117 --> 00:04:58,218
and Microsoft Foundry.

84
00:04:58,917 --> 00:05:03,477
That means large enterprise teams can
deploy these exact same agentic capabilities

85
00:05:03,857 --> 00:05:07,997
inside their existing cloud environments,
backed by their existing compliance,

86
00:05:08,378 --> 00:05:10,897
logging, and multi cloud governance
frameworks.

87
00:05:11,161 --> 00:05:14,221
So, where does this leave us on the
question of agent trust?

88
00:05:15,381 --> 00:05:19,002
I used to be pretty skeptical about
letting an agent run unsupervised overnight.

89
00:05:19,721 --> 00:05:24,261
The risk of runaway hallucination or
corrupted state was just too high.

90
00:05:24,321 --> 00:05:28,121
But with isolated worktree sandboxes,
dedicated secret vaults,

91
00:05:28,521 --> 00:05:31,462
and execution happening inside persistent
cloud environments,

92
00:05:31,941 --> 00:05:33,902
my comfort level has completely shifted.

93
00:05:33,917 --> 00:05:38,157
I can now hand off routine maintenance
chores, database audits,

94
00:05:38,197 --> 00:05:42,397
and internal dashboard builds to a digital
teammate before going to sleep,

95
00:05:42,437 --> 00:05:46,877
knowing the credentials are secure and the
execution environment is completely

96
00:05:46,941 --> 00:05:47,597
contained.

97
00:05:47,810 --> 00:05:52,877
We are finally moving past the era of
shiny chat wrappers and entering the era of

98
00:05:52,957 --> 00:05:56,077
genuine, reliable software colleagues.

99
00:05:56,402 --> 00:05:58,762
That is what has me excited this week.

100
00:05:58,782 --> 00:06:02,702
Give Routines and Secrets a spin in your
own workflow and see what you think.

101
00:06:02,762 --> 00:06:03,802
Catch you all in the next one.