CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

ainativedev/aidevcon-2026-nyc

AI Native DevCon 2026 New York — practical agent skills grounded in conference talks

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

transcript.mdskills/talk-braun-context-graphs/references/

Corrected transcript: Models Alone Aren't Enough: Engineering Context as a Graph for Human + Agent Teams

Alan Braun · AI Native DevCon New York 2026

Recording: https://www.youtube.com/watch?v=pH9CZKTB4yQ

Source: original corrected SRT supplied for this talk. Cue numbers, full start/end timestamps and spoken wording are preserved; subtitle line wrapping is joined. Inserted conference promotion cues 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.

For topic navigation, see outline.md.

Cue 1 — 00:00:00,000 --> 00:00:03,343

All right. Let's make sense of that title and I'll give you a little flavor of what

Cue 2 — 00:00:03,383 --> 00:00:08,103

I do. So at Atlassian, I run the ecosystem: I run our developer platform,

Cue 3 — 00:00:08,143 --> 00:00:12,903

I run marketplace and I run the Atlassian MCP and a bunch of API things underneath

Cue 4 — 00:00:12,943 --> 00:00:16,263

it. I would imagine a lot of you have used the Atlassian MCP.

Cue 5 — 00:00:16,303 --> 00:00:21,783

If I'm if I'm right, I have the pleasure of working with developers every day as

Cue 6 — 00:00:21,823 --> 00:00:25,583

they build on our platform, largely building on our development platform,

Cue 7 — 00:00:25,623 --> 00:00:29,863

Forge. If you don't know what that is, I encourage you to look it up or maybe talk

Cue 8 — 00:00:29,903 --> 00:00:34,503

to me about it later. Now, individually, I'm doing what probably a lot of you are

Cue 9 — 00:00:34,543 --> 00:00:38,543

doing. I have agents running all night long, coding, building something I'm

Cue 10 — 00:00:38,583 --> 00:00:43,263

personally building all the time, and we're constantly leveraging what I would call

Cue 11 — 00:00:43,303 --> 00:00:48,503

context when we do that creating CLAUDE.md files, AGENTS.md files,

Cue 12 — 00:00:48,543 --> 00:00:52,023

markdown files, trying to figure out how to get the agents to be smarter,

Cue 13 — 00:00:52,063 --> 00:00:54,443

to move faster, to deliver more.

Cue 14 — 00:00:54,943 --> 00:01:00,023

Now imagine that at scale. When you have multiple teams in an organization,

Cue 15 — 00:01:00,063 --> 00:01:02,823

multiple repositories, lots of knowledge.

Cue 16 — 00:01:02,863 --> 00:01:07,263

How do you bring that all together into what I'm going to call the connected context

Cue 17 — 00:01:07,303 --> 00:01:12,563

to help you move quicker? Now, we've heard two great talks this morning.

Cue 18 — 00:01:13,063 --> 00:01:14,823

This morning was Software Factory.

Cue 19 — 00:01:14,863 --> 00:01:18,863

I'm moving faster and the last talk was all about context I hope I'm going to build

Cue 20 — 00:01:18,903 --> 00:01:22,803

on. I was just nodding my head, going, yes, yes, yes, yes, yes.

Cue 21 — 00:01:22,863 --> 00:01:25,863

Hopefully I'm going to build on that and give you a bit of perspective.

Cue 22 — 00:01:25,903 --> 00:01:28,680

I'm going to do three things in this talk.

Cue 23 — 00:01:28,823 --> 00:01:33,583

First, I'm going to talk about why we're not quite moving faster with AI yet and

Cue 24 — 00:01:33,623 --> 00:01:35,283

what we need to do to get there.

Cue 25 — 00:01:35,503 --> 00:01:39,303

I'm going to talk about the connected graph and how the graph can help us.

Cue 26 — 00:01:39,343 --> 00:01:43,863

And then I'm actually going to show you a live demo of an engineer actually using

Cue 27 — 00:01:43,903 --> 00:01:48,163

some of the stuff I'm going to talk to to show you how it works in practice.

Cue 28 — 00:01:49,503 --> 00:01:54,203

So the first thing I think we need to ask is, is AI delivering value?

Cue 29 — 00:01:54,863 --> 00:01:57,143

Now, you meet with a lot of customers and a lot of developers,

Cue 30 — 00:01:57,183 --> 00:01:59,783

and I think we would get a lot of mixed answers here.

Cue 31 — 00:01:59,823 --> 00:02:01,543

Individually, we're crushing it.

Cue 32 — 00:02:01,583 --> 00:02:05,783

We're executing better. We're executing more, we're using agents.

Cue 33 — 00:02:05,823 --> 00:02:09,043

We're doing all the things that we're told and we're supposed to be doing.

Cue 34 — 00:02:09,303 --> 00:02:13,663

But at an organizational level, I'm not sure we would get that response.

Cue 35 — 00:02:13,703 --> 00:02:18,343

I don't think it's meeting quite the enterprise transformation that companies are

Cue 36 — 00:02:18,383 --> 00:02:23,703

expecting. In fact, to make this a bit more real, in our State of Teams report,

Cue 37 — 00:02:23,743 --> 00:02:29,323

we reported that 89% of executives say AI has increased their speed of work.

Cue 38 — 00:02:29,553 --> 00:02:35,543

Think we would all not agree to that, but only 6% feel confident that they can point

Cue 39 — 00:02:35,583 --> 00:02:40,683

to org-wide returns from AI. That is a massive disconnect.

Cue 40 — 00:02:41,063 --> 00:02:42,903

So why is the disconnect happening?

Cue 41 — 00:02:42,943 --> 00:02:48,803

And what's interesting about this is models and intelligence keep getting better.

Cue 42 — 00:02:49,663 --> 00:02:51,263

The capabilities keep going up.

Cue 43 — 00:02:51,303 --> 00:02:54,383

And I think we would all agree as we use new models like oh my God,

Cue 44 — 00:02:54,423 --> 00:02:56,243

look at what these models can do.

Cue 45 — 00:02:56,423 --> 00:02:59,583

And just to give you perspective, some of the newer models are not on this chart

Cue 46 — 00:02:59,623 --> 00:03:04,483

because we don't have benchmarks yet, but the capabilities continue to increase.

Cue 47 — 00:03:04,943 --> 00:03:09,083

What I want you to actually look at on this slide though is the orange line,

Cue 48 — 00:03:09,143 --> 00:03:13,963

which is our productivity. And our productivity seems to plateau.

Cue 49 — 00:03:14,503 --> 00:03:16,183

So what's the disconnect here.

Cue 50 — 00:03:16,223 --> 00:03:20,723

Models intelligence higher but productivity is plateauing.

Cue 51 — 00:03:21,423 --> 00:03:25,423

So why isn't AI delivering these transformations if we're getting all this

Cue 52 — 00:03:25,463 --> 00:03:28,243

intelligence what's actually the problem?

Cue 53 — 00:03:29,543 --> 00:03:35,983

At Atlassian, we like to think about it as this equation AI outcomes equals

Cue 54 — 00:03:36,023 --> 00:03:41,343

context times intelligence. I just told you we have a lot of intelligence.

Cue 55 — 00:03:41,383 --> 00:03:45,663

I think we would all agree. Think I'm falling in love with Claude Opus 5.5 every

Cue 56 — 00:03:45,703 --> 00:03:52,623

day. It's amazing what it can do, but without context, it's hard to really make it

Cue 57 — 00:03:52,663 --> 00:03:57,583

do what you needed to do in the most productive way and the fastest way.

Cue 58 — 00:03:57,623 --> 00:04:03,383

And I think something we care a lot about is at a cost efficient way with the least

Cue 59 — 00:04:03,423 --> 00:04:04,723

amount of tokens.

Cue 60 — 00:04:06,383 --> 00:04:10,183

And so before I get into context, I want to talk about what it means to be AI

Cue 61 — 00:04:10,223 --> 00:04:14,903

native. And this is something that we spend a lot of time thinking about.

Cue 62 — 00:04:14,943 --> 00:04:19,143

On the left side, you have individual productivity, one person,

Cue 63 — 00:04:19,183 --> 00:04:24,263

faster AI, faster tasks, one repository, a bunch of skills that you use,

Cue 64 — 00:04:24,303 --> 00:04:29,623

all the local stuff. And I think we would all agree we're kind of crushing it there.

Cue 65 — 00:04:29,663 --> 00:04:32,983

We're producing more, we're getting more work done.

Cue 66 — 00:04:33,023 --> 00:04:35,263

I don't know that we're necessarily shipping faster.

Cue 67 — 00:04:35,303 --> 00:04:39,483

Maybe we are in some cases, we're definitely creating a lot faster.

Cue 68 — 00:04:39,583 --> 00:04:43,343

And I think in my world, being a product person that works with my design and

Cue 69 — 00:04:43,383 --> 00:04:48,343

engineering counterparts, those lines are really blurring now because of the

Cue 70 — 00:04:48,383 --> 00:04:50,263

individual productivity that we're getting.

Cue 71 — 00:04:50,303 --> 00:04:54,743

But I want to move to the right side is the right side is really where we start to

Cue 72 — 00:04:54,783 --> 00:05:00,403

get the big gains and the big impact from AI and it's workflow productivity.

Cue 73 — 00:05:01,143 --> 00:05:05,223

How do we get more efficient with planning, with building or with reviewing and with

Cue 74 — 00:05:05,263 --> 00:05:10,623

shipping? The rest of the process where it's complicated, you need multiple teams

Cue 75 — 00:05:10,663 --> 00:05:14,503

working with teams. There's collaboration, there's a lot of things that need to

Cue 76 — 00:05:14,543 --> 00:05:20,303

happen and work right. And how can we leverage AI and end to end business context to

Cue 77 — 00:05:20,343 --> 00:05:22,523

focus on the right side of this?

Cue 78 — 00:05:23,783 --> 00:05:27,383

So I've said context. This word is overused a lot.

Cue 79 — 00:05:27,423 --> 00:05:32,343

I know we just heard that in the last talk, but what should a good context layer

Cue 80 — 00:05:32,383 --> 00:05:35,343

solve for before we talk about what it is?

Cue 81 — 00:05:35,383 --> 00:05:37,683

What's the problem we're solving?

Cue 82 — 00:05:37,983 --> 00:05:41,863

I think there's two problems that I just want to call out that if you're going to

Cue 83 — 00:05:41,903 --> 00:05:44,983

build your own context layer, you should think about these two things.

Cue 84 — 00:05:45,023 --> 00:05:47,083

Number one, work isn't siloed.

Cue 85 — 00:05:47,703 --> 00:05:50,483

It spans teams, tools and agents.

Cue 86 — 00:05:51,863 --> 00:05:55,883

So if you're going to build a context layer, it's got to work across those things.

Cue 87 — 00:05:56,823 --> 00:06:01,683

And the bottlenecks move. Let's dive into that I'll explain what I mean in a minute.

Cue 88 — 00:06:01,783 --> 00:06:04,443

First, let's talk about work for a minute.

Cue 89 — 00:06:05,223 --> 00:06:07,943

I think work is a team sport. I think we know that.

Cue 90 — 00:06:07,983 --> 00:06:12,023

I think unless you're an individual person doing a project in any organization,

Cue 91 — 00:06:12,063 --> 00:06:16,263

you're working with teams. And for years we've been collaborating together with

Cue 92 — 00:06:16,303 --> 00:06:19,643

teams. And it's not always easy to work with another team.

Cue 93 — 00:06:20,023 --> 00:06:23,383

You need to think hard about how am I going to influence that other team to do what

Cue 94 — 00:06:23,423 --> 00:06:28,423

I need to support the product that I'm working on, and that's context that we use as

Cue 95 — 00:06:28,463 --> 00:06:32,723

humans to figure out how do we get those teams to work together.

Cue 96 — 00:06:32,863 --> 00:06:35,163

But now what's happening is something bigger.

Cue 97 — 00:06:35,503 --> 00:06:37,643

We're bringing agents into the mix.

Cue 98 — 00:06:37,823 --> 00:06:40,203

So now agents are working with teams.

Cue 99 — 00:06:40,783 --> 00:06:44,383

But this is a problem because the agents don't always have the same context the

Cue 100 — 00:06:44,423 --> 00:06:48,523

human has, or even worse, they have more context.

Cue 101 — 00:06:48,663 --> 00:06:51,723

And now how do they work efficiently with other teams?

Cue 102 — 00:06:51,983 --> 00:06:57,023

Let's take it a step further. Now we have agents to agents and agents are working

Cue 103 — 00:06:57,063 --> 00:07:03,463

with other agents. How do we build a connected context that solves for all of these

Cue 104 — 00:07:03,503 --> 00:07:08,483

use cases teams to teams, teams to agents and agents to agents.

Cue 105 — 00:07:09,143 --> 00:07:14,423

The other problem we have here is that the context the agent has access to needs to

Cue 106 — 00:07:14,463 --> 00:07:18,563

have the same permissions that you do as a human in your team.

Cue 107 — 00:07:18,703 --> 00:07:23,303

What happens if the agent has access to more information than they're supposed to,

Cue 108 — 00:07:23,343 --> 00:07:25,783

and now they bring it into that team environment?

Cue 109 — 00:07:25,823 --> 00:07:30,103

This can get pretty scary. So we have a lot to think about as we think about what

Cue 110 — 00:07:30,143 --> 00:07:34,763

are we building for? Well, let's also talk about bottlenecks for a minute.

Cue 111 — 00:07:35,423 --> 00:07:38,783

Historically, I think we all would agree, especially in this room,

Cue 112 — 00:07:38,823 --> 00:07:42,403

that building, coding and implementing was the long pole in the tent.

Cue 113 — 00:07:42,983 --> 00:07:46,943

It took a lot of time to go execute planning probably should have taken longer,

Cue 114 — 00:07:46,983 --> 00:07:48,983

but I don't know that we all planned very well.

Cue 115 — 00:07:49,023 --> 00:07:51,583

So we went into code and we executed a lot.

Cue 116 — 00:07:51,623 --> 00:07:54,403

And then we did review testing and all that, but we ship.

Cue 117 — 00:07:54,823 --> 00:07:57,343

Okay, so that's the typical of what we've seen in the past.

Cue 118 — 00:07:57,383 --> 00:07:59,143

Guess what? I think we all know this.

Cue 119 — 00:07:59,183 --> 00:08:02,623

We're executing faster. We're delivering more PRs.

Cue 120 — 00:08:02,663 --> 00:08:06,183

If you look at our studies with our product, I will tell you,

Cue 121 — 00:08:06,223 --> 00:08:12,543

we see PRs are on the rise and we're approaching near 100% of PRs are actually AI

Cue 122 — 00:08:12,583 --> 00:08:15,123

driven. So we're executing really quickly.

Cue 123 — 00:08:15,303 --> 00:08:17,843

That means we're releasing faster, right?

Cue 124 — 00:08:17,903 --> 00:08:20,103

That is the reality. We're executing faster.

Cue 125 — 00:08:20,143 --> 00:08:22,083

How are we not releasing quicker?

Cue 126 — 00:08:22,823 --> 00:08:24,623

Well, here's the problem. I think we all smile.

Cue 127 — 00:08:24,663 --> 00:08:26,583

I see a couple laughs in the audience.

Cue 128 — 00:08:26,623 --> 00:08:31,203

It's because we have this tax that we've put on the organization now.

Cue 129 — 00:08:31,583 --> 00:08:35,663

We're definitely executing and writing more code, but that's putting a lot more

Cue 130 — 00:08:35,703 --> 00:08:38,563

pressure on the back half of the process.

Cue 131 — 00:08:38,823 --> 00:08:44,423

So what actually happens is this it's more PRs, it's more approvals,

Cue 132 — 00:08:44,463 --> 00:08:47,783

it's more decisions, it's more sign offs, it's more meetings.

Cue 133 — 00:08:47,823 --> 00:08:52,383

And I keep asking myself, how do I eliminate the meetings at the end of the process?

Cue 134 — 00:08:52,423 --> 00:08:55,143

What can I do to simplify that down?

Cue 135 — 00:08:55,183 --> 00:08:58,203

Because let me tell you guys, this is the opportunity.

Cue 136 — 00:08:58,863 --> 00:09:02,863

It's not necessarily the front part, although I think you can do things up front

Cue 137 — 00:09:02,903 --> 00:09:06,183

which fix the back half. How do you fix the back?

Cue 138 — 00:09:06,223 --> 00:09:12,203

And a good layer optimizes for every stage, not just execution.

Cue 139 — 00:09:12,383 --> 00:09:17,143

And I think historically when we talk about context, we've been very focused on how

Cue 140 — 00:09:17,183 --> 00:09:21,523

do we execute faster, not necessarily the entire workflow.

Cue 141 — 00:09:22,143 --> 00:09:24,423

So I've said this word a lot context.

Cue 142 — 00:09:24,463 --> 00:09:28,343

And I know in the previous talk I loved it because it was very aligned to the way I

Cue 143 — 00:09:28,383 --> 00:09:32,383

think about context. I want to give you a sense, just so your mental model is the

Cue 144 — 00:09:32,423 --> 00:09:35,903

way I'm thinking from here going forward as I present this out.

Cue 145 — 00:09:35,943 --> 00:09:39,363

This is a graph and this is how I think about context.

Cue 146 — 00:09:40,023 --> 00:09:44,423

It starts with a project or some piece of work and the people are connected around

Cue 147 — 00:09:44,463 --> 00:09:47,043

it. Connected is the key word.

Cue 148 — 00:09:47,183 --> 00:09:49,023

What people are working on that project?

Cue 149 — 00:09:49,063 --> 00:09:52,223

What are their roles? Where are they in the organization?

Cue 150 — 00:09:52,263 --> 00:09:55,983

What do they know? We then have knowledge connected to that project.

Cue 151 — 00:09:56,023 --> 00:10:00,983

Knowledge can come in. Our world could be Confluence, but it could be Google Docs,

Cue 152 — 00:10:01,023 --> 00:10:05,623

it could be markdown files. It's all the stuff that we know about that project.

Cue 153 — 00:10:05,663 --> 00:10:08,663

And then I can go on and on. What's the goals of the organization?

Cue 154 — 00:10:08,703 --> 00:10:11,303

What's the business metrics? What's the sales data?

Cue 155 — 00:10:11,343 --> 00:10:15,143

What's the customer context? What's the whiteboards that we've done?

Cue 156 — 00:10:15,183 --> 00:10:20,943

What's the meetings? All of this needs to be connected in order to actually

Cue 157 — 00:10:20,983 --> 00:10:26,503

streamline the workflow. And so when we connect all of this,

Cue 158 — 00:10:26,543 --> 00:10:30,563

that's when we get the opportunity to cut down the review cycles.

Cue 159 — 00:10:30,823 --> 00:10:35,903

And that's when we get the opportunity to really shorten the entire process and

Cue 160 — 00:10:35,943 --> 00:10:41,243

actually end up with faster delivery, not just faster execution.

Cue 161 — 00:10:42,103 --> 00:10:46,263

So where does all this context live though, is the problem, and how do I take

Cue 162 — 00:10:46,303 --> 00:10:52,223

advantage of it? Well, in our world, in an Atlassian world, we have something called

Cue 163 — 00:10:52,263 --> 00:10:54,903

a Teamwork Graph, which I'll talk about in a minute, which connects this.

Cue 164 — 00:10:54,943 --> 00:10:58,563

But in a lot of people's world, you already have this stuff at the edge.

Cue 165 — 00:10:58,743 --> 00:11:00,343

This is stuff you're probably doing already.

Cue 166 — 00:11:00,383 --> 00:11:02,783

Let's talk through it. Your instruction files.

Cue 167 — 00:11:02,823 --> 00:11:07,583

You're already using Claude and agents, I'm sure, and you're getting your agents in

Cue 168 — 00:11:07,623 --> 00:11:09,543

some cases to make them smarter.

Cue 169 — 00:11:09,583 --> 00:11:13,543

I hope in other cases you're actually using human intelligence to go do it,

Cue 170 — 00:11:13,583 --> 00:11:16,403

because I think we get a lot of slop in those files.

Cue 171 — 00:11:16,663 --> 00:11:21,583

We have specs. Something that we're learning at Atlassian is spending a lot of time

Cue 172 — 00:11:21,623 --> 00:11:26,263

up front with the spec can reduce a lot of time in the back half.

Cue 173 — 00:11:26,303 --> 00:11:30,503

Reviewing the code, reviewing the PR, and actually spending time shipping.

Cue 174 — 00:11:30,543 --> 00:11:34,703

So we are now in a process of a lot of time up front in the specs.

Cue 175 — 00:11:34,743 --> 00:11:38,543

And I think if we heard this morning with Tessl in the software factory,

Cue 176 — 00:11:38,583 --> 00:11:41,863

I think it's a great example of really getting it right up front so it can go

Cue 177 — 00:11:41,903 --> 00:11:47,063

through the process. Then we have skills and what are the important things that

Cue 178 — 00:11:47,103 --> 00:11:49,983

skills have to do is trim that context.

Cue 179 — 00:11:50,023 --> 00:11:53,503

Part of the problem is we bloat the context window and we need to be really careful.

Cue 180 — 00:11:53,543 --> 00:11:55,463

We have the right things at the right time.

Cue 181 — 00:11:55,503 --> 00:11:59,743

Our code index I think you know all the repository is connected and memory.

Cue 182 — 00:11:59,783 --> 00:12:01,023

We heard this in the last talk.

Cue 183 — 00:12:01,063 --> 00:12:04,543

It's how do I not make the same mistakes over and over and over?

Cue 184 — 00:12:04,583 --> 00:12:09,083

How do I learn to get faster and smarter as I go through the process?

Cue 185 — 00:12:09,383 --> 00:12:13,783

And then there's certainly MCP servers where we can connect things using tools.

Cue 186 — 00:12:13,823 --> 00:12:16,683

We can go out to third party services and connect them.

Cue 187 — 00:12:16,943 --> 00:12:21,723

But what ultimately brings this together is when you have something that's org wide,

Cue 188 — 00:12:21,863 --> 00:12:25,823

that takes all of these sources and it connects them together.

Cue 189 — 00:12:25,863 --> 00:12:30,183

That's the opportunity here. So why org wide context?

Cue 190 — 00:12:30,223 --> 00:12:32,743

Why does that matter? Why do I care about this.

Cue 191 — 00:12:32,783 --> 00:12:36,903

Well I'm going to ask three questions that I would say are pretty important as you

Cue 192 — 00:12:36,943 --> 00:12:41,763

think about this. Going back to your desk is why is this code like this?

Cue 193 — 00:12:42,343 --> 00:12:48,143

It's a really hard question to answer if you don't have all the context.

Cue 194 — 00:12:48,183 --> 00:12:52,343

You go from the code to the pull request, maybe to the work item to understand why

Cue 195 — 00:12:52,383 --> 00:12:57,343

it was done, and then to the documentation or the decision log of why things happen,

Cue 196 — 00:12:57,383 --> 00:13:01,463

the way they happened, or who should review this change.

Cue 197 — 00:13:01,503 --> 00:13:04,663

It seems so primitive, but it's so hard.

Cue 198 — 00:13:04,703 --> 00:13:08,983

How do you connect the changes happening in the organization to the only team,

Cue 199 — 00:13:09,023 --> 00:13:13,003

to the roles in the team? And by the way, it's not just coders.

Cue 200 — 00:13:13,103 --> 00:13:16,703

A lot of times I have to go talk to legal, to finance, to marketing,

Cue 201 — 00:13:16,743 --> 00:13:21,303

to a lot of other organizations that I got to figure out what they're working on.

Cue 202 — 00:13:21,343 --> 00:13:25,703

And I often find there's a lot of stuff written in marketing that would have

Cue 203 — 00:13:25,743 --> 00:13:28,703

informed the way that we're actually developing the product,

Cue 204 — 00:13:28,743 --> 00:13:32,943

and we missed that. If it's not connected, what breaks?

Cue 205 — 00:13:32,983 --> 00:13:37,023

If I change this, this is another one that's really hard.

Cue 206 — 00:13:37,063 --> 00:13:39,303

You're working on a service or a product.

Cue 207 — 00:13:39,343 --> 00:13:44,103

You have dependent services, you have their teams, and you have open incidents and

Cue 208 — 00:13:44,143 --> 00:13:49,303

customer requests coming in. How do I get visibility into this to get smarter,

Cue 209 — 00:13:49,343 --> 00:13:52,323

to make me more productive as I release?

Cue 210 — 00:13:53,503 --> 00:13:55,663

For Atlassian, the answer is Teamwork Graph.

Cue 211 — 00:13:55,703 --> 00:13:57,823

That's what we use. We use it internally.

Cue 212 — 00:13:57,863 --> 00:13:59,923

I'm going to show you an example of that.

Cue 213 — 00:14:00,383 --> 00:14:02,623

We also ship this if you're using our products.

Cue 214 — 00:14:02,663 --> 00:14:05,623

This is over 20 years of us working.

Cue 215 — 00:14:05,663 --> 00:14:10,383

Understanding how people do work, understanding work teams, goals and knowledge and

Cue 216 — 00:14:10,423 --> 00:14:15,543

how it comes together. And we have a connected graph for our products as well as

Cue 217 — 00:14:15,583 --> 00:14:20,343

connectors in the industry to bring in outside, which allows you to take advantage

Cue 218 — 00:14:20,383 --> 00:14:22,563

of the graph to get your work done.

Cue 219 — 00:14:22,783 --> 00:14:27,343

Now there's a lot of talk about with Teamwork Graph, but what I'm more interested in

Cue 220 — 00:14:27,383 --> 00:14:29,063

giving you what's underneath it.

Cue 221 — 00:14:29,103 --> 00:14:32,743

Just in case you don't want to use Teamwork Graph, you want to do your own thing.

Cue 222 — 00:14:32,783 --> 00:14:36,063

There are seven layers of context that we think about.

Cue 223 — 00:14:36,103 --> 00:14:40,083

First, this knowledge. It's everything your team has written down and learned.

Cue 224 — 00:14:40,423 --> 00:14:44,063

Next is communication. It's all the conversations, the meetings,

Cue 225 — 00:14:44,103 --> 00:14:47,943

the text chats, the emails. The Slacks, the Teams.

Cue 226 — 00:14:47,983 --> 00:14:50,343

Next is the work. Where do you track your work?

Cue 227 — 00:14:50,383 --> 00:14:54,603

So in our case it's Jira. But where do you actually track your work?

Cue 228 — 00:14:54,663 --> 00:14:59,163

The code? How do we connect all the code repositories into one place?

Cue 229 — 00:14:59,343 --> 00:15:03,063

The people, which is so undervalued and underestimated.

Cue 230 — 00:15:03,103 --> 00:15:05,343

It's the heart of your organization.

Cue 231 — 00:15:05,383 --> 00:15:08,023

It's where people sit, where they work their teams.

Cue 232 — 00:15:08,063 --> 00:15:11,543

And again, going back to that marketing example, how do I connect to all the

Cue 233 — 00:15:11,583 --> 00:15:16,943

organizations, the assets? I know there was an automotive example used this morning.

Cue 234 — 00:15:16,983 --> 00:15:21,383

We have the same thing where we see automotive companies are writing software

Cue 235 — 00:15:21,423 --> 00:15:25,423

connected to hardware. How do you connect both of those when you're making

Cue 236 — 00:15:25,463 --> 00:15:29,383

decisions? And then the data and the data is all the business data.

Cue 237 — 00:15:29,423 --> 00:15:31,943

It's the Databricks, the Snowflakes, the data lakes.

Cue 238 — 00:15:31,983 --> 00:15:35,203

How do I connect it all together to make decisions?

Cue 239 — 00:15:35,303 --> 00:15:38,623

Now where this shows up is in a couple different places.

Cue 240 — 00:15:38,663 --> 00:15:42,463

Okay. For us, there's over 100 powered experiences you can have.

Cue 241 — 00:15:42,503 --> 00:15:46,223

I'm going to break this down into three areas, which I think we often don't think

Cue 242 — 00:15:46,263 --> 00:15:51,263

about in terms of context. When I say context in this room, I would imagine most of

Cue 243 — 00:15:51,303 --> 00:15:55,263

you immediately think about your coding agents and how can I bring context to my

Cue 244 — 00:15:55,303 --> 00:15:59,223

coding agent. There's two other places where I think it's really important to think

Cue 245 — 00:15:59,263 --> 00:16:04,583

about chat. How do I chat in an organization and make sure I get the right

Cue 246 — 00:16:04,623 --> 00:16:09,303

organizational context to get my response and search?

Cue 247 — 00:16:09,343 --> 00:16:13,103

How do I find the right things at the right time inside of my organization?

Cue 248 — 00:16:13,143 --> 00:16:17,223

This context graph helps solve for all three of the use cases.

Cue 249 — 00:16:17,263 --> 00:16:22,043

So yes, certainly agents writing code, but also search and also chat.

Cue 250 — 00:16:22,943 --> 00:16:28,223

Now how do we access the the Teamwork Graph or how do you access your context layer?

Cue 251 — 00:16:28,263 --> 00:16:30,303

There are two ways that we do it today.

Cue 252 — 00:16:30,343 --> 00:16:32,143

A lot of you use the MCP server.

Cue 253 — 00:16:32,183 --> 00:16:36,263

I'm not sure any of you even know or have used the Teamwork Graph CLI,

Cue 254 — 00:16:36,303 --> 00:16:37,783

but there are multiple use cases.

Cue 255 — 00:16:37,823 --> 00:16:40,103

Depending on the use case, the CLI can be better.

Cue 256 — 00:16:40,143 --> 00:16:43,623

Sometimes MCP can be better, other times.

Cue 257 — 00:16:43,663 --> 00:16:47,563

The reason why you bring this up is I want to talk about MCP for a minute.

Cue 258 — 00:16:47,623 --> 00:16:52,143

Some of you may be building MCPs and some of you might be building if you build a

Cue 259 — 00:16:52,183 --> 00:16:56,843

context graph, or if you build a context layer, you may want to build an MCP server

Cue 260 — 00:16:56,883 --> 00:16:59,883

and you should. That's connected to that context.

Cue 261 — 00:17:00,183 --> 00:17:03,663

One of the biggest learnings that we've learned is actually two that I'm going to

Cue 262 — 00:17:03,703 --> 00:17:08,433

give you with MCP. Number one, you have to be really careful with tool bloat.

Cue 263 — 00:17:08,823 --> 00:17:13,623

So in that world, I own the MCP across the platform, which is every one of our

Cue 264 — 00:17:13,663 --> 00:17:16,183

products. You can imagine what happened.

Cue 265 — 00:17:16,223 --> 00:17:18,763

Every product team created their own tools.

Cue 266 — 00:17:19,183 --> 00:17:20,863

We threw the tools altogether.

Cue 267 — 00:17:20,903 --> 00:17:25,743

And in the early days of MCP, it meant loading all those tools into the context

Cue 268 — 00:17:25,783 --> 00:17:29,383

window we've now optimized. We now search for the right tool,

Cue 269 — 00:17:29,423 --> 00:17:32,903

but the reality is people don't want product tools.

Cue 270 — 00:17:32,943 --> 00:17:36,063

They want platform tools, cross-product tools.

Cue 271 — 00:17:36,103 --> 00:17:39,603

What they really want you to think about is the jobs to be done.

Cue 272 — 00:17:40,263 --> 00:17:45,063

And so what I encourage you to think about is working backwards from the jobs to be

Cue 273 — 00:17:45,103 --> 00:17:49,263

done and thinking about what are the right tools that you should build for your

Cue 274 — 00:17:49,303 --> 00:17:54,983

context layer. At the end of the day, what's so important is that the agent gets the

Cue 275 — 00:17:55,023 --> 00:17:58,503

right context at the right time to do the job.

Cue 276 — 00:17:58,543 --> 00:18:02,423

That is, doing more context doesn't mean better here.

Cue 277 — 00:18:02,463 --> 00:18:05,023

The right context means better.

Cue 278 — 00:18:05,063 --> 00:18:09,543

So really thinking through how do you access your context layer.

Cue 279 — 00:18:09,583 --> 00:18:13,283

And there's a little byproduct to this, which I think is interesting.

Cue 280 — 00:18:13,383 --> 00:18:15,743

The last talk we talked about MCP connectors.

Cue 281 — 00:18:15,783 --> 00:18:17,483

And by the way you should use them.

Cue 282 — 00:18:17,703 --> 00:18:21,423

But what's interesting about Teamwork Graph is we use the connectors to get the

Cue 283 — 00:18:21,463 --> 00:18:27,543

context, so we can get the right references for what we need for this task.

Cue 284 — 00:18:27,583 --> 00:18:33,143

And then we can use the right MCP connector to go get the details of that reference

Cue 285 — 00:18:33,183 --> 00:18:35,443

at the right time to do the job.

Cue 286 — 00:18:35,503 --> 00:18:38,023

So we're not using MCP connectors for everything.

Cue 287 — 00:18:38,063 --> 00:18:42,703

We're using it for the right things that we need, and that our CLI has over 670

Cue 288 — 00:18:42,743 --> 00:18:44,683

commands that you can use also.

Cue 300 — 00:19:26,707 --> 00:19:29,993

So to bring this to life, I'm going to show you a video.

Cue 301 — 00:19:30,033 --> 00:19:34,753

The video runs slightly long, but I want you to get a sense of a real engineer doing

Cue 302 — 00:19:34,793 --> 00:19:36,873

real things with this context layer.

Cue 303 — 00:19:36,913 --> 00:19:41,313

What you're going to witness is a Claude Code on the left without context,

Cue 304 — 00:19:41,353 --> 00:19:43,553

a Claude Code on the right with context.

Cue 305 — 00:19:43,593 --> 00:19:46,653

And you're going to be able to see how they perform differently.

Cue 306 — 00:19:46,797 --> 00:19:51,677

Hey, folks, it's me here. Helping us explore a good way to demonstrate how the

Cue 307 — 00:19:51,717 --> 00:19:55,617

context of our Teamwork Graph helps agents do our work better.

Cue 308 — 00:19:55,957 --> 00:19:59,157

I'm demonstrating this with two Claude Code side by side.

Cue 309 — 00:19:59,197 --> 00:20:02,197

On the left, we have a freshly installed Claude Code.

Cue 310 — 00:20:02,237 --> 00:20:06,697

On the right we have exactly the same Claude Code but with skills installed.

Cue 311 — 00:20:06,877 --> 00:20:11,917

So I'll show you if I type skills here, we'll see that there are only three skills.

Cue 312 — 00:20:11,957 --> 00:20:14,077

These are built in from Claude Code.

Cue 313 — 00:20:14,117 --> 00:20:17,237

If we do the same thing on the right, we'll see.

Cue 314 — 00:20:17,277 --> 00:20:20,737

There are a bunch of skills in addition to those three.

Cue 315 — 00:20:20,877 --> 00:20:26,357

These skills will teach Claude Code how to use our CLI to fetch context from our

Cue 316 — 00:20:26,397 --> 00:20:29,957

Teamwork Graph. That is the only difference between these two versions.

Cue 317 — 00:20:29,997 --> 00:20:36,077

So we're going to run the same tasks both in the same Pi platform.

Cue 318 — 00:20:36,117 --> 00:20:40,857

This is our platform integration repo and I will paste the prompt here.

Cue 319 — 00:20:40,997 --> 00:20:47,357

So the same prompt this prompt this task is a real task that's about extending a

Cue 320 — 00:20:47,397 --> 00:20:51,797

stream hub pipeline to support some new events while keeping compatibility with

Cue 321 — 00:20:51,837 --> 00:20:56,557

relevant services. It's a non-trivial feature to add, so we're going to ask the

Cue 322 — 00:20:56,597 --> 00:21:00,757

agents to just come up with a plan and not the implementation,

Cue 323 — 00:21:00,797 --> 00:21:03,477

just the technical plan. So we'll let them run.

Cue 324 — 00:21:03,517 --> 00:21:06,797

Now it's going to take a while quite a while.

Cue 325 — 00:21:06,837 --> 00:21:12,437

And I know it's it's not going to be easy to parse this whole like these two screens

Cue 326 — 00:21:12,477 --> 00:21:17,517

from two scrolling terminals. So I captured the full trajectories from those two

Cue 327 — 00:21:17,557 --> 00:21:22,037

agent runs and visualize them here to help us see the differences more easily.

Cue 328 — 00:21:22,077 --> 00:21:28,997

The prompt we just sent was really about improving a Jira admin's Slack digest to

Cue 329 — 00:21:29,037 --> 00:21:33,497

remove already approved access requests from the pending list.

Cue 330 — 00:21:33,597 --> 00:21:38,357

We're watching the agents working in real time here, and we'll monitor how the

Cue 331 — 00:21:38,397 --> 00:21:42,477

cumulative token cost and time spans and the outcome is.

Cue 332 — 00:21:42,517 --> 00:21:47,677

Here we can start to see, with TWG available here on the inside.

Cue 333 — 00:21:47,717 --> 00:21:53,517

This agent, the first couple of turns was really mostly about searching for context

Cue 334 — 00:21:53,557 --> 00:21:57,317

through search. And it's searching for context.

Cue 335 — 00:21:57,357 --> 00:22:02,197

It found some Confluence pages and it learned some of the rights of the for what you

Cue 336 — 00:22:02,237 --> 00:22:07,137

search next. Here I can show quickly what the pages are.

Cue 337 — 00:22:07,277 --> 00:22:11,717

These two pages are the pages that got found and the agents read these pages.

Cue 338 — 00:22:11,757 --> 00:22:17,677

These pages largely just explained how the Slack Jira integration works and what the

Cue 339 — 00:22:17,717 --> 00:22:21,717

access request notification is, really how they work, what they do,

Cue 340 — 00:22:21,757 --> 00:22:24,877

and the product behavior is here as well.

Cue 341 — 00:22:24,917 --> 00:22:28,677

So even humans, we will find this context very helpful.

Cue 342 — 00:22:28,717 --> 00:22:34,517

And then the agents used a code search to identify some other related files.

Cue 343 — 00:22:34,557 --> 00:22:37,317

So here we can see some of these results.

Cue 344 — 00:22:37,357 --> 00:22:40,277

Can't even be found through typical lexical search.

Cue 345 — 00:22:40,317 --> 00:22:43,357

It was the semantic search index that found these matches.

Cue 346 — 00:22:43,397 --> 00:22:46,617

Because you can see it's not exactly match the keywords.

Cue 347 — 00:22:47,237 --> 00:22:49,957

So that is something that was quite useful as well.

Cue 348 — 00:22:49,997 --> 00:22:54,637

The raw Claude Code on the left without TWG is just like jumping directly into the

Cue 349 — 00:22:54,677 --> 00:22:57,217

local repo and starting to read the files.

Cue 350 — 00:22:57,597 --> 00:23:02,397

I'll speed through this. But the important thing is really through this search

Cue 351 — 00:23:02,437 --> 00:23:07,357

operations, we actually found there is another repo that is a dependency of what we

Cue 352 — 00:23:07,397 --> 00:23:11,477

are about to change here. And in this other repo, there is some very important

Cue 353 — 00:23:11,517 --> 00:23:15,457

context that can anchor the implementation that we need to do here.

Cue 354 — 00:23:15,997 --> 00:23:19,297

So yeah, let's speed through this.

Cue 355 — 00:23:19,877 --> 00:23:25,157

And here we can see with TWG, the agent was actually able to produce the correct

Cue 356 — 00:23:25,197 --> 00:23:30,577

plan sooner and with less cost, while the raw Claude Code is still running,

Cue 357 — 00:23:30,717 --> 00:23:36,297

eventually the raw Claude Code will also finish.

Cue 358 — 00:23:36,357 --> 00:23:43,217

And the problem is that it lacked context and missed the related repo in its plan.

Cue 359 — 00:23:43,397 --> 00:23:46,197

And as a result, the plan was actually is wrong.

Cue 360 — 00:23:46,237 --> 00:23:52,317

It hallucinated the event schema and would actually send a duplicate notification to

Cue 361 — 00:23:52,357 --> 00:23:54,657

the users, which will be really bad.

Cue 362 — 00:23:54,757 --> 00:23:59,957

On the other side, we have a correct plan and that's because it got informed by the

Cue 363 — 00:23:59,997 --> 00:24:04,517

information, the context that exists in this other repo, which is really hard to

Cue 364 — 00:24:04,557 --> 00:24:08,857

find unless we do the search and content retrieval up front.

Cue 365 — 00:24:09,277 --> 00:24:14,917

So Claude Code equipped with TWG successfully identified this hidden context,

Cue 366 — 00:24:14,957 --> 00:24:19,717

and made the correct plan with less cost and less time.

Cue 367 — 00:24:19,757 --> 00:24:21,797

Yeah. And that is the power of context.

Cue 368 — 00:24:21,837 --> 00:24:26,717

The agent with TWG spends the first couple of turns mostly on context retrieval,

Cue 369 — 00:24:26,757 --> 00:24:31,997

before going into the local repos to do the work, and the upfront context found

Cue 370 — 00:24:32,037 --> 00:24:35,482

through TWG was really the key to its success here.

Cue 371 — 00:24:37,756 --> 00:24:41,276

Okay, so I think through the example, it really brings it to life of what happened

Cue 372 — 00:24:41,316 --> 00:24:43,396

on the left side versus the right side.

Cue 373 — 00:24:43,436 --> 00:24:46,316

On the left side, it was trying to do the more traditional approach,

Cue 374 — 00:24:46,356 --> 00:24:50,476

using MCP servers, going out and getting the context it thought it needed,

Cue 375 — 00:24:50,516 --> 00:24:53,596

and on the right side it uses the context graph.

Cue 376 — 00:24:53,636 --> 00:24:57,676

But what's really important is the upfront work, as mentioned,

Cue 377 — 00:24:57,716 --> 00:25:01,616

of getting that context. Now I want to give you a bit of perspective here.

Cue 378 — 00:25:01,676 --> 00:25:07,356

What we've seen in our tests is we spend up to 48% fewer tokens and

Cue 379 — 00:25:07,396 --> 00:25:13,476

44% better answer quality. So if they're like if there's a slide that I would say to

Cue 380 — 00:25:13,516 --> 00:25:17,236

land on here for connected context, it's here.

Cue 381 — 00:25:17,276 --> 00:25:19,156

This is the reason why it matters.

Cue 382 — 00:25:19,196 --> 00:25:23,236

Your answers are better, less hallucination at a lower cost.

Cue 383 — 00:25:23,276 --> 00:25:28,036

And so just working through that, I think that's the really the reason why this

Cue 384 — 00:25:28,076 --> 00:25:32,996

matters. So either way what is a graph do and what are the things I want you to

Cue 385 — 00:25:33,036 --> 00:25:36,916

think about. And I'm going to wrap with just giving you some takeaways of what you

Cue 386 — 00:25:36,956 --> 00:25:39,836

can go do. If you're using Teamwork Graph, that's great.

Cue 387 — 00:25:39,876 --> 00:25:43,156

If you're not, what are other things that you can do in your own environment to take

Cue 388 — 00:25:43,196 --> 00:25:47,256

advantage today? So one is connect every source.

Cue 389 — 00:25:48,036 --> 00:25:52,796

This is where it starts. And you can use MCPs to to start this.

Cue 390 — 00:25:52,836 --> 00:25:57,356

But thinking through all the sources that you need to actually get your work done

Cue 391 — 00:25:57,396 --> 00:25:59,716

faster and better is a great start.

Cue 392 — 00:25:59,756 --> 00:26:04,556

Number two is link your work. Think about putting issue keys in PRs.

Cue 393 — 00:26:04,596 --> 00:26:06,396

You might be doing some of this already.

Cue 394 — 00:26:06,436 --> 00:26:12,196

Even that alone will actually connect the work back to the issue or the PR that

Cue 395 — 00:26:12,236 --> 00:26:15,396

you're actually working on. Owners code owners.

Cue 396 — 00:26:15,436 --> 00:26:19,116

If you're not doing this, you should be, but getting ownership of who owns what.

Cue 397 — 00:26:19,156 --> 00:26:24,176

And remember, that's great for repositories, but maybe not organizational ownership.

Cue 398 — 00:26:24,516 --> 00:26:29,056

The fourth one is really important to trim the context to what matters.

Cue 399 — 00:26:29,716 --> 00:26:34,556

And number five is protect because permissions are super important here.

Cue 400 — 00:26:34,596 --> 00:26:38,836

If we're not careful, agents get access to things they shouldn't have access to.

Cue 401 — 00:26:38,876 --> 00:26:41,696

So you always need to be thinking about protection.

Cue 402 — 00:26:41,916 --> 00:26:44,116

Now don't take our word for it.

Cue 403 — 00:26:44,156 --> 00:26:45,876

You need to test this. You need to measure it.

Cue 404 — 00:26:45,916 --> 00:26:49,796

And I know in the last we talked about this also you got to do a bunch of evals.

Cue 405 — 00:26:49,836 --> 00:26:52,396

And I just want to give you how we think about this.

Cue 406 — 00:26:52,436 --> 00:26:58,476

Take five prompts, write the prompts of what you do today and actually write out

Cue 407 — 00:26:58,516 --> 00:27:00,576

what you expect the outcome to be.

Cue 408 — 00:27:00,756 --> 00:27:05,636

You then pass it through just like you saw with context, with your current context,

Cue 409 — 00:27:05,676 --> 00:27:10,236

with your connected context. And then you use a blind LLM to judge it.

Cue 410 — 00:27:10,276 --> 00:27:13,064

Don't tell the LLM which one you're going through.

Cue 411 — 00:27:13,104 --> 00:27:16,836

Just have the LLM score the results, okay?

Cue 412 — 00:27:16,876 --> 00:27:20,996

And you look at what's better, and you have to constantly fine tune this.

Cue 413 — 00:27:21,036 --> 00:27:25,916

You have to continually iterate, get a better and better to get your context to

Cue 414 — 00:27:25,956 --> 00:27:27,896

actually deliver what you need.

Cue 415 — 00:27:28,396 --> 00:27:30,736

And so I'm going to end on this slide.

Cue 416 — 00:27:31,076 --> 00:27:32,836

You already have most of this.

Cue 417 — 00:27:32,876 --> 00:27:35,596

This isn't necessarily new. It already lives in your environment.

Cue 418 — 00:27:35,636 --> 00:27:40,396

The work is connecting them. And it's helping give your agents one step to actually

Cue 419 — 00:27:40,436 --> 00:27:44,796

use this information. So as you go back to your desk, there's there's really six

Cue 420 — 00:27:44,836 --> 00:27:47,356

things I want you to think about one.

Cue 421 — 00:27:47,396 --> 00:27:51,036

Link your work. If nothing else, whatever system you're using,

Cue 422 — 00:27:51,076 --> 00:27:55,136

if it's Jira, great, if not whatever you're using, put your issue keys inside of

Cue 423 — 00:27:55,176 --> 00:28:00,056

your PRs and make sure as part of your CI it actually checks for it.

Cue 424 — 00:28:00,316 --> 00:28:02,716

Number two is write down ownership.

Cue 425 — 00:28:02,756 --> 00:28:04,956

And I would go beyond just code ownership.

Cue 426 — 00:28:04,996 --> 00:28:08,476

I would actually think about organizational ownership somewhere where the agent can

Cue 427 — 00:28:08,516 --> 00:28:12,416

look up who are the people associated with the work happening.

Cue 428 — 00:28:12,796 --> 00:28:15,536

Number three is write down your decisions.

Cue 429 — 00:28:16,196 --> 00:28:19,956

The agent can do a lot of this work for you, but it's so important that you're what

Cue 430 — 00:28:19,996 --> 00:28:23,516

actually happens. So we can learn from it and we can leverage it.

Cue 431 — 00:28:23,556 --> 00:28:26,616

So the next task can use that context.

Cue 432 — 00:28:26,996 --> 00:28:29,436

Number four connect 3 to 5 sources.

Cue 433 — 00:28:29,476 --> 00:28:33,356

Don't do 15. Connect 3 to 5. Start small.

Cue 434 — 00:28:33,396 --> 00:28:35,436

Continue to build up. See what matters.

Cue 435 — 00:28:35,476 --> 00:28:39,896

Continue to run evals and see how do I actually get better results.

Cue 436 — 00:28:40,636 --> 00:28:45,436

Write one context skill a skill that knows how to use all of the stuff that you've

Cue 437 — 00:28:45,476 --> 00:28:50,596

written down trims. It has the right permissions and passes in the right context for

Cue 438 — 00:28:50,636 --> 00:28:55,156

the right moment. Going back to what I said earlier, it's not just teams and teams,

Cue 439 — 00:28:55,196 --> 00:28:57,956

it's teams and agents. It's agents and agents.

Cue 440 — 00:28:57,996 --> 00:29:03,516

How do we give it the right context and run your five prompt test before and after,

Cue 441 — 00:29:03,556 --> 00:29:08,116

over and over and over. I will end where I started.

Cue 442 — 00:29:08,156 --> 00:29:14,836

Models are super intelligent, but the reality is more context doesn't actually mean

Cue 443 — 00:29:14,876 --> 00:29:20,396

it's better. It's connected context that actually fuels those models to become more

Cue 444 — 00:29:20,436 --> 00:29:23,336

intelligent and allows you to deliver faster.

Cue 445 — 00:29:23,596 --> 00:29:26,056

Thank you. Hope you have a great conference.

skills

tile.json