WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=9_oy4owGtPI

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/9_oy4owGtPI.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:13.040 --> 00:00:14.320
Good to meet you folks.

s3
00:00:14.320 --> 00:00:16.240
Thanks for everyone who's here.

s4
00:00:16.240 --> 00:00:20.160
Um, today we're going to talk about sharing context, right?

s5
00:00:20.160 --> 00:00:22.960
Everything in AI is about the data you give it.

s6
00:00:22.960 --> 00:00:25.199
Bad data, bad outputs, right?

s7
00:00:25.199 --> 00:00:29.439
Agent memory is a solved problem, but agent learning is not.

s8
00:00:29.439 --> 00:00:37.760
And we're going to talk about not just like you know what we're building etc. It's about how why we built it and how we're using it for ourselves.

s9
00:00:37.760 --> 00:00:38.719
Okay.

s10
00:00:38.719 --> 00:00:50.000
So as we said agent memory works you know there's a bunch of stuff you feed it conversations etc. And you're able to extract memories from the conversation and agents are able to reuse them.

s11
00:00:50.000 --> 00:00:51.680
It's per agent.

s12
00:00:51.680 --> 00:00:55.280
However learning is never done in isolation.

s13
00:00:55.280 --> 00:01:03.039
It's when a group of agents come together and need to work with each other, they have to now start learning and sharing, right?

s14
00:01:03.039 --> 00:01:16.400
And that is an unsolved problem because we're infrastructure people and infrastructure built for a single agent to do memory is often not the same infrastructure that is needed in order to learn, right?

s15
00:01:16.400 --> 00:01:19.360
We're going to talk a little bit more about that.

s16
00:01:19.360 --> 00:01:27.280
Okay, so by way of background, I'm Karthik, one of the co-founders and a CEO of a database company, Yugabyte DB, right?

s17
00:01:27.280 --> 00:01:28.640
So what the hell are we doing here?

s18
00:01:28.640 --> 00:01:32.560
Why are we building a aentic data infrastructure?

s19
00:01:32.560 --> 00:01:33.680
What are we doing doing that?

s20
00:01:33.680 --> 00:01:34.159
Right?

s21
00:01:34.159 --> 00:01:36.320
We actually found this the hard way.

s22
00:01:36.320 --> 00:01:37.920
We run agents for ourselves.

s23
00:01:37.920 --> 00:01:40.799
We run two agents internal to the company.

s24
00:01:40.799 --> 00:01:46.799
Hagen for customer support and growth vector for a coordinated go to market between sales and marketing.

s25
00:01:46.799 --> 00:01:50.640
And we run two agents externally for our customers and users.

s26
00:01:50.640 --> 00:01:55.040
The first one is Perf Advisor, which is an agentic database admin, right?

s27
00:01:55.040 --> 00:01:58.240
So it runs databases for our customers as a database company.

s28
00:01:58.240 --> 00:01:59.360
Pretty useful.

s29
00:01:59.360 --> 00:02:05.040
And Voyager, which helps them migrate and modernize from other databases to a cloudnative one, right?

s30
00:02:05.040 --> 00:02:06.880
So four agents.

s31
00:02:06.880 --> 00:02:10.160
We failed to ship anything and make it work at all.

s32
00:02:10.160 --> 00:02:11.039
It was really hard.

s33
00:02:11.039 --> 00:02:28.640
We've been trying for almost a year before we realized we were failing at the data problem right and if a database company is failing at a data problem be it for agentic it becomes personal right so yes it became personal we're like what the hell's going on and when we then realized that a lot of people were

s34
00:02:28.640 --> 00:02:41.760
asking us like yes we can do scalable vectors we can do scalable graph we can do scalable relational or people came to us and said more scale more performance less cost we're like wait a minute at some point this is ridiculous right it's breaking the laws physics.

s35
00:02:41.760 --> 00:02:44.319
So what is the real failure?

s36
00:02:44.319 --> 00:02:44.800
Right?

s37
00:02:44.800 --> 00:02:55.920
When you cut everything o apart and just go to the core, it's because agent A, my agent is doing some work and I need to work with say one of you guys, say maybe Heather, right?

s38
00:02:55.920 --> 00:03:00.800
My who's going to be showing a demo and I am not able to give that context.

s39
00:03:00.800 --> 00:03:03.360
I'm only able to give the output.

s40
00:03:03.360 --> 00:03:08.959
The output being an MD file, a PDF, a document, a something.

s41
00:03:08.959 --> 00:03:11.200
But I'm not able to give the reasoning behind it.

s42
00:03:11.200 --> 00:03:11.680
Right?

s43
00:03:11.680 --> 00:03:12.879
So what does this lose?

s44
00:03:12.879 --> 00:03:14.959
This loses reasoning.

s45
00:03:14.959 --> 00:03:19.680
This loses the deadends, the number of failures you've had before you actually succeeded.

s46
00:03:19.680 --> 00:03:21.360
And it loses the whole context.

s47
00:03:21.360 --> 00:03:22.000
Right?

s48
00:03:22.000 --> 00:03:26.480
That means Heather's agent is going to be rederiving all of this stuff.

s49
00:03:26.480 --> 00:03:27.519
And it's not for free.

s50
00:03:27.519 --> 00:03:30.879
You're burning tokens and paying somebody else for it.

s51
00:03:30.879 --> 00:03:33.840
This is the same as a group of humans working together.

s52
00:03:33.840 --> 00:03:42.560
When a new person joins the team, they're going to be told, "We're going to do X." And they come up with 50 reasons why we should do something else.

s53
00:03:42.560 --> 00:03:46.239
And each time, yes, we considered that two weeks ago, two months ago.

s54
00:03:46.239 --> 00:03:47.840
That's not how humans operate.

s55
00:03:47.840 --> 00:03:51.200
We actually bring the person in and says, "Okay, look, this is the context.

s56
00:03:51.200 --> 00:03:52.640
This is the history of the project.

s57
00:03:52.640 --> 00:03:53.760
This is how it's done.

s58
00:03:53.760 --> 00:03:55.120
Now, let's go from here."

s59
00:03:55.120 --> 00:03:58.640
That context is missing for agents.

s60
00:03:58.720 --> 00:04:00.959
So, five things, misconceptions.

s61
00:04:00.959 --> 00:04:02.400
Okay, really quick.

s62
00:04:02.400 --> 00:04:03.439
More memory.

s63
00:04:03.439 --> 00:04:05.680
Agentic memory is not learning.

s64
00:04:05.680 --> 00:04:06.720
It's just more memory, right?

s65
00:04:06.720 --> 00:04:08.400
Like it's more expensive.

s66
00:04:08.400 --> 00:04:11.280
Shared state is not shared knowledge.

s67
00:04:11.280 --> 00:04:17.600
Just because you can share something with somebody else, like an MD file or something, doesn't mean you have shared knowledge.

s68
00:04:17.600 --> 00:04:22.639
Rag transcripts, let me rag everything, is not equal to learning and quality.

s69
00:04:22.639 --> 00:04:29.520
It just means there's a whole bunch of garbage there and now LLMs get to burn even more tokens to give you even worse results.

s70
00:04:29.520 --> 00:04:31.840
Fine-tuning is not scalable.

s71
00:04:31.840 --> 00:04:39.280
You cannot, you know, save and persist lessons that agents have come up with through fine-tuning and building custom infrastructure.

s72
00:04:39.280 --> 00:04:44.639
And a bigger context window is the shest way to quickest failure in an expensive manner, right?

s73
00:04:44.639 --> 00:04:45.600
Because it doesn't do anything.

s74
00:04:45.600 --> 00:04:47.199
Your context rots.

s75
00:04:47.199 --> 00:04:51.040
Every single one of these is a architecture gap, right?

s76
00:04:51.040 --> 00:04:56.080
So this is what we realized we needed to plug for ourselves, right?

s77
00:04:56.080 --> 00:04:56.240
Right.

s78
00:04:56.240 --> 00:05:03.199
And we've been finding a lot of users are are really enjoying and really liking this approach of thinking about context.

s79
00:05:03.199 --> 00:05:09.280
Firstly, if there's a context that agent A can share with agent B, it closes the lost context problem.

s80
00:05:09.280 --> 00:05:12.479
They're working on the same context.

s81
00:05:12.479 --> 00:05:15.280
Second thing is this context can degrade.

s82
00:05:15.280 --> 00:05:18.560
That means not everybody can contribute to that context.

s83
00:05:18.560 --> 00:05:19.759
It's just like code.

s84
00:05:19.759 --> 00:05:21.360
Not anyone can contribute.

s85
00:05:21.360 --> 00:05:24.880
You need to go through a process of supervised promotion.

s86
00:05:24.880 --> 00:05:27.039
And only then does it make it durable.

s87
00:05:27.039 --> 00:05:32.160
And lastly, you need to be able to ask your the question, where the hell did this piece of information come from?

s88
00:05:32.160 --> 00:05:33.600
It makes no sense at all.

s89
00:05:33.600 --> 00:05:40.320
And be able to trace it back to something, a data source, a person, a conversation from which it was extracted.

s90
00:05:40.320 --> 00:05:42.880
And that's the first step to improving your context.

s91
00:05:42.880 --> 00:05:49.440
So we started building this by assuming how hard can it be.

s92
00:05:49.440 --> 00:05:52.240
So the first thing we did was it's just a knowledge base.

s93
00:05:52.240 --> 00:05:53.680
We throw everything in the knowledge base.

s94
00:05:53.680 --> 00:05:55.039
is going to be great, right?

s95
00:05:55.039 --> 00:05:56.320
And how wrong we were.

s96
00:05:56.320 --> 00:06:01.680
We built a whole distributed rag pipeline, extremely scalable, very elastic, amazing.

s97
00:06:01.680 --> 00:06:02.720
It was just amazing, right?

s98
00:06:02.720 --> 00:06:04.479
And it didn't do anything good for us.

s99
00:06:04.479 --> 00:06:05.840
And then we said, you know what?

s100
00:06:05.840 --> 00:06:07.520
This whole shared stuff doesn't work.

s101
00:06:07.520 --> 00:06:09.680
We'll let every agent spin up a tiny database.

s102
00:06:09.680 --> 00:06:10.639
We're a database company.

s103
00:06:10.639 --> 00:06:11.520
How hard can it be?

s104
00:06:11.520 --> 00:06:13.520
Tiny Postgress, make it efficient.

s105
00:06:13.520 --> 00:06:19.280
Well, we realized that every repeatable workflow was being reinvented on the fly.

s106
00:06:19.280 --> 00:06:20.400
Didn't work at all.

s107
00:06:20.400 --> 00:06:21.840
And then we said, "Okay, fine.

s108
00:06:21.840 --> 00:06:31.280
we'll just have some backend that does a whole bunch of processing and you know we'll just you know not even bother with the UI cuz agents don't need UI and then we realized we couldn't tell what was going on.

s109
00:06:31.280 --> 00:06:34.880
No human in the loop equals really bad results.

s110
00:06:34.880 --> 00:06:35.600
Okay.

s111
00:06:35.600 --> 00:06:38.319
And then we thought okay at least we have the rag pipeline.

s112
00:06:38.319 --> 00:06:49.840
Well, it turned out even that was shitty because, you know, we took all the best components from the, you know, whatever the academia academia says and we put it all in there and we got poor numbers.

s113
00:06:49.840 --> 00:06:59.280
14% faithfulness using an MD benchmark, public benchmark, 20% on a PDF, 57% context precision using all of this other stuff.

s114
00:06:59.280 --> 00:07:07.120
And when we actually started looking and tracing and doing our traces and tuning and experimenting, we realized a lot of things were off.

s115
00:07:07.120 --> 00:07:13.039
And as we improved them one variable at a time, we were able to jump to 65 74 82.

s116
00:07:13.039 --> 00:07:13.360
Right?

s117
00:07:13.360 --> 00:07:15.039
And here's the kicker.

s118
00:07:15.039 --> 00:07:23.120
Our context chunks, which is the context size, went down from over 7,000 chunks to about a,000 chunks.

s119
00:07:23.120 --> 00:07:31.199
So 17th the context size, a lot higher precision and a lot less money burned on tokens, right?

s120
00:07:31.199 --> 00:07:34.160
So that's where Mako comes in.

s121
00:07:34.160 --> 00:07:35.039
What is Mako?

s122
00:07:35.039 --> 00:07:41.840
It's a agentnative persistence layer that takes all the mistakes we made and all the things we got right and puts it into one box.

s123
00:07:41.840 --> 00:07:42.560
Right?

s124
00:07:42.560 --> 00:07:46.479
And the way Mako works, it exposes a data pack, right?

s125
00:07:46.479 --> 00:07:51.199
As a database company, we don't call it a database because it has multiple types of databases.

s126
00:07:51.199 --> 00:07:54.080
Vector, graph, relational, no SQL, whatever, right?

s127
00:07:54.080 --> 00:08:03.360
And what it does is stores aspects of the context, knowledge base, memory, conversations, observability, a lot more, right?

s128
00:08:03.360 --> 00:08:14.479
And what it enables is it enables agents to come in and push pieces of their context into a private space and promote that to a shared space, right?

s129
00:08:14.479 --> 00:08:28.000
And then other agents or humans that come in can now go against that promoted highquality context and quickly retrieve stuff and you can now continue to add more people and more agents and do the handoff efficiently.

s130
00:08:28.000 --> 00:08:28.639
Right?

s131
00:08:28.639 --> 00:08:30.080
Hopefully that makes sense.

s132
00:08:30.080 --> 00:08:37.279
Um the idea is take the durable repeatable workflows and make it infrastructure.

s133
00:08:37.279 --> 00:08:42.560
Okay, before we go into a demo, right, a confession.

s134
00:08:42.560 --> 00:08:49.120
This slide deck that you're looking at was built using Makeo as the context persistence layer, right?

s135
00:08:49.120 --> 00:08:59.920
The way it worked was Heather, who's going to give you show you the demo, actually wrote up an outline and created a data pack in Mako and simply sh sent sent me a share request,

s136
00:08:59.920 --> 00:09:00.399
right?

s137
00:09:00.399 --> 00:09:02.320
And she said, "I'm doing the demo.

s138
00:09:02.320 --> 00:09:03.680
Here's the outline of my demo.

s139
00:09:03.680 --> 00:09:07.600
Here's the rough talk track that needs to lead up to it because that's what makes sense."

s140
00:09:07.600 --> 00:09:10.800
I took it and I'm I'm sure she used some agent harness.

s141
00:09:10.800 --> 00:09:11.519
I don't know what it is.

s142
00:09:11.519 --> 00:09:12.080
I really don't know.

s143
00:09:12.080 --> 00:09:13.040
I didn't ask her.

s144
00:09:13.040 --> 00:09:14.880
I didn't even have a meeting with her.

s145
00:09:14.880 --> 00:09:15.920
She just sent me this.

s146
00:09:15.920 --> 00:09:21.360
I plugged it into my cloud desktop and said, "Take this outline and break it down for me."

s147
00:09:21.360 --> 00:09:22.800
And it gave me a few bullet points.

s148
00:09:22.800 --> 00:09:24.240
This is what you're talking about.

s149
00:09:24.240 --> 00:09:25.200
I'm like, "Okay, great.

s150
00:09:25.200 --> 00:09:26.800
I want to change a couple of things."

s151
00:09:26.800 --> 00:09:29.680
Changed a couple of things and said, "Here's the new outline.

s152
00:09:29.680 --> 00:09:31.839
Can you build me my slide deck based on this?"

s153
00:09:31.839 --> 00:09:32.800
And it did.

s154
00:09:32.800 --> 00:09:36.560
And that diagram on the left, it built a really crappy diagram, right?

s155
00:09:36.560 --> 00:09:37.519
It didn't look good at all.

s156
00:09:37.519 --> 00:09:38.720
So, I said, like, you know what?

s157
00:09:38.720 --> 00:09:43.200
Everyone's saying this uh cloud design is the rage, so why don't I give that thing a go?

s158
00:09:43.200 --> 00:09:44.800
And it actually did a better job.

s159
00:09:44.800 --> 00:09:45.760
So, you know, there you go.

s160
00:09:45.760 --> 00:09:47.440
That image came from cloud design.

s161
00:09:47.440 --> 00:09:49.279
Everything else came from cloud desktop.

s162
00:09:49.279 --> 00:09:56.240
Shot it back to Heather and, you know, Heather took the final pass over it to get hook it up with the demo and here we are.

s163
00:09:56.240 --> 00:09:58.560
So, we barely talked, right?

s164
00:09:58.560 --> 00:09:59.279
We didn't need to.

s165
00:09:59.279 --> 00:10:05.519
We just shared context and our agents did the work and uh you know, you can go you can watch it.

s166
00:10:05.519 --> 00:10:08.959
I think maybe I'll invite Heather to talk to this slide and and the thing after.

s167
00:10:08.959 --> 00:10:11.279
I hope the demo's good because I haven't seen it myself.

s168
00:10:11.279 --> 00:10:13.519
So, yeah.

s169
00:10:20.640 --> 00:10:25.120
The idea is to watch collective knowledge grow over time, right?

s170
00:10:25.120 --> 00:10:29.519
What Nico provides uh that I was able to use for this is a resumability.

s171
00:10:29.519 --> 00:10:37.680
So resumability I'm going to show you is being able to start a session and then kill it and then start it again with the same agent.

s172
00:10:37.680 --> 00:10:42.000
Right now if you do it with something like cloud desktop, they have projects for this, right?

s173
00:10:42.000 --> 00:10:42.240
Right?

s174
00:10:42.240 --> 00:10:46.000
So if you do a cloud project, everything's together or if you turn on memory.

s175
00:10:46.000 --> 00:10:55.920
But in this case, I guess like two months ago, I turned off all the memory on all of my idees because it doesn't matter what if I use something while I'm on the go, like a mobile app.

s176
00:10:55.920 --> 00:11:00.560
Maybe I use chat EBT that isn't as sophisticated, right, as some of my coding idees.

s177
00:11:00.560 --> 00:11:03.920
And so maybe I want to continue working on that topic.

s178
00:11:03.920 --> 00:11:08.079
I can do that because all of the context is shared in our cloud now.

s179
00:11:08.079 --> 00:11:25.040
And by sharing I mean that obviously chatbt is a different um agent but it's still my account so I should be able to get to it right but because we're security focused and security minded we don't want like just any agent to get to it so I have to promote the kind of memory that that agent can

s180
00:11:25.040 --> 00:11:36.160
also retrieve in like a collective pool and because you do this the cost itself goes down tremendously because the first time I figure stuff out I'm going back and forth with my LLM

s181
00:11:36.160 --> 00:11:42.560
I'm burning lots of tokens to get to an end goal, whether that is code or it's just a knowledge work.

s182
00:11:42.560 --> 00:11:48.160
But the next time that I want to retrieve it from another agent or another session, it doesn't need to burn through that.

s183
00:11:48.160 --> 00:11:53.519
It just calls to the Miko memory and then it brings it right back with far less token.

s184
00:11:53.519 --> 00:11:56.959
So, we're already saving a lot of token spend.

s185
00:11:56.959 --> 00:12:09.040
The alternative to this would be to just again keep things in empty files or keep things where you have to constantly traverse with every single agent's LLM and burn through tokens again and again in order to retrieve.

s186
00:12:09.040 --> 00:12:10.000
Does that make sense?

s187
00:12:10.000 --> 00:12:13.440
The difference here is just an MCP tool call.

s188
00:12:13.440 --> 00:12:23.839
So to kind of prove that out because we have sensational internet here, I have recorded this for you at a little bit of a faster speed hopefully.

s189
00:12:23.839 --> 00:12:24.240
All right.

s190
00:12:24.240 --> 00:12:25.839
See this is resumability.

s191
00:12:25.839 --> 00:12:32.160
So here I'm saying that I want to store um the conversation in a certain data pack, right?

s192
00:12:32.160 --> 00:12:36.320
Let's call it the early access launch and I want to put a piece of information in it.

s193
00:12:36.320 --> 00:12:39.200
In this case, um our signup emails failed.

s194
00:12:39.200 --> 00:12:44.160
True story because we actually only used the SCS's sandbox instead of the prod.

s195
00:12:44.160 --> 00:12:47.839
So uh we kind of went over the limit there when we were testing things.

s196
00:12:47.839 --> 00:12:53.279
And that's a good piece of information for maybe our our ops team, right?

s197
00:12:53.279 --> 00:13:05.120
So in this process, it's calling data pack create uh because I've connected the MCP server to clot here and I've created a skill that will tell it what to do.

s198
00:13:05.120 --> 00:13:10.399
You can decide how often you want it to call or if you just want to do it judiciously here.

s199
00:13:10.399 --> 00:13:15.279
And so now it's storing the incident fact as a searchable memory with memory add.

s200
00:13:15.279 --> 00:13:16.480
And then it's done now.

s201
00:13:16.480 --> 00:13:18.800
So the data pack has been created.

s202
00:13:18.800 --> 00:13:23.360
there's a conversation in there and it'll be retrievable in the in the future sessions.

s203
00:13:23.360 --> 00:13:24.800
That sounds great.

s204
00:13:24.800 --> 00:13:25.519
Awesome.

s205
00:13:25.519 --> 00:13:28.959
Now I've killed that session and notice I'm not in projects.

s206
00:13:28.959 --> 00:13:31.040
This is just a new chat window.

s207
00:13:31.040 --> 00:13:35.519
Now I'm going to ask the same question so I can resume.

s208
00:13:35.519 --> 00:13:42.480
Before I had to again have to make sure I turn on cloud memory or turn or put it inside of a project, but it's not in either one of these.

s209
00:13:42.480 --> 00:13:48.079
I could be logged in with the same account in like claw.ai, which would be technically a different agent.

s210
00:13:48.079 --> 00:13:53.839
So I'm trying to prove that you can at least resume your sessions after maybe days with the same agent.

s211
00:13:53.839 --> 00:14:03.519
So the fail result there was actually that it tried to search first instead of creating uh a conversation a new conversation here.

s212
00:14:03.519 --> 00:14:05.440
We want to save everyone that we do.

s213
00:14:05.440 --> 00:14:11.920
So it's the MCP server instructing to create a new conversation for it so we can store this.

s214
00:14:11.920 --> 00:14:20.720
And see now they searched Mo and it found that hey this is what happened back in May and it was able to pull things back out.

s215
00:14:22.000 --> 00:14:24.480
All right for sharability.

s216
00:14:24.959 --> 00:14:26.560
Let's go ahead and get this done.

s217
00:14:26.560 --> 00:14:29.199
If you notice this is a completely different IDE.

s218
00:14:29.199 --> 00:14:31.279
In this case it's codeex.

s219
00:14:31.279 --> 00:14:38.079
Um so I've asked it questions like okay um do you have any known cause for the signup emails that are failing?

s220
00:14:38.079 --> 00:14:43.279
And so it's gonna look and it's gonna look and it says, "Nope, I don't see anything.

s221
00:14:43.279 --> 00:14:44.160
I don't see anything.

s222
00:14:44.160 --> 00:14:45.600
I'm looking."

s223
00:14:45.600 --> 00:14:54.639
Um, it did say that it saw that there was a brand new uh data pack that it's connected to, but it can't actually see anything.

s224
00:14:54.639 --> 00:14:55.760
And why is that?

s225
00:14:55.760 --> 00:15:02.240
Because I actually haven't shared that private memory from my claw desktop yet uh into a shared space.

s226
00:15:02.240 --> 00:15:04.959
So right now, those memories are still private.

s227
00:15:04.959 --> 00:15:12.079
If you notice the early access data pack is seen because I have connected it but I just haven't promoted it yet.

s228
00:15:12.079 --> 00:15:15.519
So in that case I'm going to go to sign out.

s229
00:15:15.519 --> 00:15:24.560
I'm going to sign in with my other account here to show you that yes I do have access to this data pack and I need to promote that piece of memory

s230
00:15:24.560 --> 00:15:27.040
into a shared knowledge space.

s231
00:15:27.040 --> 00:15:34.480
Allow it to go ahead and and go through and then I it says it doesn't see it but I said well same question.

s232
00:15:34.480 --> 00:15:37.040
Let's try it again.

s233
00:15:37.920 --> 00:15:44.520
So, now that I've promoted it into that space, sped up for your pleasure here

s234
00:15:44.520 --> 00:15:44.959
[laughter]

s235
00:15:44.959 --> 00:15:47.920
because we all know how long-winded our LLMs can get.

s236
00:15:47.920 --> 00:15:51.279
Um, now it can find that piece of memory.

s237
00:15:51.279 --> 00:15:54.320
It can find the piece of memory because I've allowed that one piece to go.

s238
00:15:54.320 --> 00:15:55.440
It's very fine grained.

s239
00:15:55.440 --> 00:15:56.720
It doesn't have to be.

s240
00:15:56.720 --> 00:16:09.199
You can open the whole data pack if you want as a maintainer to it, but this is much more ideal if you're just trying to give bespoke pieces of information, maybe best practices, maybe something that is the ongoing

s241
00:16:09.199 --> 00:16:13.759
process that you want to keep going.

s242
00:16:14.160 --> 00:16:16.720
And then like the last part of this is the cost.

s243
00:16:16.720 --> 00:16:20.639
How do you even know that it saves you tokens?

s244
00:16:20.639 --> 00:16:30.720
If you go into your data pack, you can look at the conversations that are stored in there and how they were done because we give you embedded language traces for free in this.

s245
00:16:30.720 --> 00:16:39.600
And you can actually look at the decision trace all the way down to the tool call and prove whether or not it called as much and burned as many tokens.

s246
00:16:43.920 --> 00:16:45.920
Right?

s247
00:16:48.880 --> 00:16:54.880
So, human promotions today kind of equal the training signal for self-improving orchestration tomorrow.

s248
00:16:54.880 --> 00:16:59.440
Every human that says this is worth keeping is kind of a labelled example of what good promotion looks like.

s249
00:16:59.440 --> 00:17:06.480
So, you can actually start training maybe in the future an agent that's trained on your taste and your preferences long term.

s250
00:17:06.480 --> 00:17:09.679
But right now, we're keeping the human in the loop.

s251
00:17:09.679 --> 00:17:13.919
Those calls are just a signal, but this is actually how it starts to learn.

s252
00:17:13.919 --> 00:17:29.200
The biggest lift that I saw is that when something goes wrong, maybe in a support situation, it's hard to constantly have calls call the same server that's down all the time and then it's tails all this time.

s253
00:17:29.200 --> 00:17:30.799
Hey, do you know if this is down?

s254
00:17:30.799 --> 00:17:31.600
Is that down?

s255
00:17:31.600 --> 00:17:32.799
Slack starts blowing up.

s256
00:17:32.799 --> 00:17:34.080
How do your agents handle that?

s257
00:17:34.080 --> 00:17:37.280
They're all also wasting tokens calling the same thing.

s258
00:17:37.280 --> 00:17:46.559
If they had access to the data pack, the first agent would hit it and then tell everybody that it happened and all the other agents would know that it already happened.

s259
00:17:52.720 --> 00:17:53.039
Yeah.

s260
00:17:53.039 --> 00:17:55.120
So, yeah.

s261
00:17:55.120 --> 00:17:56.559
So, thank you, Heather.

s262
00:17:56.559 --> 00:17:58.400
Like, I think you know, pretty awesome demo.

s263
00:17:58.400 --> 00:18:06.799
I think uh if you folks were following along what happened, Heather was was in charge of the early access and launch of Mako, right?

s264
00:18:06.799 --> 00:18:20.880
and obviously used a data pack which is Mako to launch Mak Makeo right recursively and uh you know the first few people had a lot of trouble accessing their data packs on Mako and you know we you know she created a data pack for troubleshooting

s265
00:18:20.880 --> 00:18:35.200
issues with Mako a different data pack and so that's the thing that she was talking about uh a lot of us in the company you know it's shared with us so we're all able to look at it load it into our agents and able to iterate with it the folks that are in engineering they load it straight

s266
00:18:35.200 --> 00:18:50.960
into their you know cloud cloud code or codeex or like you know whatever the coding agents are like I I myself more of a cloud desktop kind of user I use it for other things like you know some a user that wanted to get access couldn't get access apology email like go through cloud help me design it

s267
00:18:50.960 --> 00:19:05.600
have a I have a data pack actually that is my voice of how I speak and same thing for posting on LinkedIn that thing can get tiresome if you want to keep writing it in your voice you know sometimes training a data pack for the context to do so will help so a lot of these things get

s268
00:19:05.600 --> 00:19:21.039
automated straight away and it gets you know makes it much much quicker to assemble right and obviously if a new person comes into the team very easy to share this context with them have them get up to speed have them contribute to it either willingly or unwillingly right sometimes they're just doing things they don't even know and

s269
00:19:21.039 --> 00:19:34.559
you extract patterns out of it oh this is really great we should make this a thing right so so like um yeah so I think these are some of the things anyways I think the bottom line is we're trying to get it closer and closer to how humans work which is how can you capture tribal knowledge how

s270
00:19:34.559 --> 00:19:39.679
can you share that How can you give controls over sharing and how can you actually tell what's going on?

s271
00:19:39.679 --> 00:19:40.000
Right.

s272
00:19:40.000 --> 00:19:40.320
Yeah.

s273
00:19:40.320 --> 00:19:41.039
Um Yeah.

s274
00:19:41.039 --> 00:19:41.440
So

s275
00:19:41.440 --> 00:19:43.600
you should you should try it for free.

s276
00:19:43.600 --> 00:19:44.320
If you want.

s277
00:19:44.320 --> 00:19:44.559
Yeah.

s278
00:19:44.559 --> 00:19:45.360
Try it for free.

s279
00:19:45.360 --> 00:19:45.600
Yeah.

s280
00:19:45.600 --> 00:19:51.200
I think the left side is you know the website makeata.ai which will take you to a login.

s281
00:19:51.200 --> 00:19:52.320
There's a free tier.

s282
00:19:52.320 --> 00:19:57.840
You can sign up and just hook it up to one of your agents and you know get your agents to start learning right right away.

s283
00:19:57.840 --> 00:19:59.200
And the right side is Discord.

s284
00:19:59.200 --> 00:20:00.960
We'd love to hear from you if you do try it out.

s285
00:20:00.960 --> 00:20:02.320
You know do drop us a note.

s286
00:20:02.320 --> 00:20:03.600
We're working on improvements.

s287
00:20:03.600 --> 00:20:05.360
That's how we communicate what we're building.

s288
00:20:05.360 --> 00:20:07.200
So do join us on Discord, right?

s289
00:20:07.200 --> 00:20:08.720
We'd love to see you guys there.

s290
00:20:08.720 --> 00:20:10.240
So I think that brings us Yeah.

s291
00:20:10.240 --> 00:20:10.799
Thank you.

s292
00:20:10.799 --> 00:20:11.120
Thanks.

s293
00:20:11.120 --> 00:20:13.440
Yeah, if we have Yeah, we're right on time actually.

s294
00:20:13.440 --> 00:20:13.760
Yeah.

s295
00:20:13.760 --> 00:20:14.960
So yeah, thank you.

s296
00:20:14.960 --> 00:20:17.600
If you guys have questions, please feel free to connect with us on the site.

s297
00:20:17.600 --> 00:20:19.679
Thank you.

s298
00:20:33.379 --> 00:20:35.379
[music]