AI Native DevCon 2026 New York — practical agent skills grounded in conference talks
86%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Bartosz Ocytko · AI Native DevCon New York 2026
Recording: https://www.youtube.com/watch?v=niMX5QHChlA
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 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.
For topic navigation, see outline.md.
Cue 1 — 00:00:00,000 --> 00:00:05,040
Thank you. So I will be sharing some of the lessons learned that we had in the
Cue 2 — 00:00:05,080 --> 00:00:07,960
company on scaling agentic engineering at Zalando.
Cue 3 — 00:00:08,000 --> 00:00:09,880
You might ask yourself what Zalando is.
Cue 4 — 00:00:09,920 --> 00:00:12,360
You may know us from our open source presence.
Cue 5 — 00:00:12,400 --> 00:00:18,440
We are the creators of Patroni, Postgres Operator and the Fashion-MNIST data set
Cue 6 — 00:00:18,480 --> 00:00:21,960
that if you have been trying ML, you might have encountered on some of the
Cue 7 — 00:00:22,000 --> 00:00:26,960
platforms, from a point of view of the business that we are doing,
Cue 8 — 00:00:27,000 --> 00:00:30,560
we are the leading European technology platform for fashion and lifestyle.
Cue 9 — 00:00:30,600 --> 00:00:35,840
We have more than 60 million customers to 150 engineering teams and 2000 developers.
Cue 10 — 00:00:35,880 --> 00:00:39,820
And on top of that come product managers, designers and the managers.
Cue 11 — 00:00:39,960 --> 00:00:44,160
We are a AWS shop. We deploy our software on Kubernetes.
Cue 12 — 00:00:44,200 --> 00:00:49,180
We love microservices in various different languages and we have a lot of them,
Cue 13 — 00:00:49,480 --> 00:00:54,680
right. So what do we build? We build a B2C and B2B platform for merchants and brands
Cue 14 — 00:00:54,720 --> 00:00:57,760
to onboard products, and we sell them to our customers.
Cue 15 — 00:00:57,800 --> 00:01:01,920
Our journey with agentic engineering started when ChatGPT came out,
Cue 16 — 00:01:01,960 --> 00:01:07,160
and our leadership made the decision to invest in a presence of an assistant on our
Cue 17 — 00:01:07,200 --> 00:01:11,440
pages, right. The assistant is basically helping our customers to discover products
Cue 18 — 00:01:11,480 --> 00:01:16,400
and to shop in different ways than just clicking through the sites from the 90s with
Cue 19 — 00:01:16,440 --> 00:01:21,440
the product detail pages that we have enhanced with AI generated videos.
Cue 20 — 00:01:21,480 --> 00:01:26,280
And our assistant has been used this year by more than 22 million customers.
Cue 21 — 00:01:26,320 --> 00:01:31,120
When we gave the team that worked on the assistant the access to the OpenAI API,
Cue 22 — 00:01:31,160 --> 00:01:35,120
we also wanted to give our developers those new tools to experiment what other use
Cue 23 — 00:01:35,160 --> 00:01:37,360
cases they would be able to build, right.
Cue 24 — 00:01:37,400 --> 00:01:41,880
So we gave them a proxy and access gated through this proxy.
Cue 25 — 00:01:41,920 --> 00:01:43,720
What is the reasoning behind that?
Cue 26 — 00:01:43,760 --> 00:01:46,640
Well, it's a central entry point for our builders.
Cue 27 — 00:01:46,680 --> 00:01:49,760
And the central entry point provides us with vendor independence.
Cue 28 — 00:01:49,800 --> 00:01:54,440
The ability to add new models from different providers and for developers is great
Cue 29 — 00:01:54,480 --> 00:01:58,480
because they just configure it in the tool that they find and open source tool.
Cue 30 — 00:01:58,520 --> 00:02:03,480
And as long as it's compliant for use at Zalando, then basically they can use the
Cue 31 — 00:02:03,520 --> 00:02:07,000
tool, which is also great as we get full visibility since the launch,
Cue 32 — 00:02:07,040 --> 00:02:10,280
and we have three years of data when it comes to model adoption,
Cue 33 — 00:02:10,320 --> 00:02:14,680
model switches. And we understand how the groups of users are switching between
Cue 34 — 00:02:14,720 --> 00:02:19,400
those models. And this also a point where we control the tool sprawl from the point
Cue 35 — 00:02:19,440 --> 00:02:22,920
of view of the technology. We settled on LiteLLM in the early days.
Cue 36 — 00:02:22,960 --> 00:02:26,560
There wasn't much choice, and we kind of have a love and hate relationship with this
Cue 37 — 00:02:26,600 --> 00:02:29,520
tool, right? We love it because of the extensibility.
Cue 38 — 00:02:29,560 --> 00:02:33,040
We have hooks. Those hooks give us anonymized user statistics.
Cue 39 — 00:02:33,080 --> 00:02:35,160
We are in Germany, so we care for privacy.
Cue 40 — 00:02:35,200 --> 00:02:38,080
We only log model information and the cost information.
Cue 41 — 00:02:38,120 --> 00:02:40,880
And through that we can calculate the statistics.
Cue 42 — 00:02:40,920 --> 00:02:43,960
The only important other thing that we log is the user agent,
Cue 43 — 00:02:44,000 --> 00:02:48,240
the header that is basically telling us what harness people are using as long as the
Cue 44 — 00:02:48,280 --> 00:02:50,280
harness is identifying itself.
Cue 45 — 00:02:50,320 --> 00:02:55,360
We can also do great things with client upgrades because users do not upgrade their
Cue 46 — 00:02:55,400 --> 00:02:58,760
harness typically, so they wait until they are forced upgraded.
Cue 47 — 00:02:58,800 --> 00:03:03,120
And even if the tool has an auto update functionality, we see that users still use
Cue 48 — 00:03:03,160 --> 00:03:08,360
for 30 or 60 days. The old versions and the team that is maintaining the proxy and
Cue 49 — 00:03:08,400 --> 00:03:11,120
then gets to deal with all the bug reports.
Cue 50 — 00:03:11,160 --> 00:03:15,040
They like to use those first upgrades to basically re baseline all teams on the new
Cue 51 — 00:03:15,080 --> 00:03:18,920
version. We also like hooks for backporting fixes, right.
Cue 52 — 00:03:18,960 --> 00:03:22,880
So if LiteLLM releases a new version and we need the bug fix immediately,
Cue 53 — 00:03:22,920 --> 00:03:27,980
then it's easier for us to backport the fix into a hook system instead of,
Cue 54 — 00:03:28,880 --> 00:03:32,240
upgrading the version. And you might ask yourself why?
Cue 55 — 00:03:32,280 --> 00:03:36,560
Well, since we added persistence with LiteLLM, there have been lots of stability
Cue 56 — 00:03:36,600 --> 00:03:40,680
issues. Those stability issues come from database migrations that are released by
Cue 57 — 00:03:40,720 --> 00:03:43,780
it. And they are not always going on cleanly.
Cue 58 — 00:03:43,840 --> 00:03:47,240
And the advice that I have for you is update often enough, because many of the bug
Cue 59 — 00:03:47,280 --> 00:03:50,200
fixes that we discover typically are fixed the newer versions,
Cue 60 — 00:03:50,240 --> 00:03:53,760
but don't stay on the bleeding edge and it will save you from the supply chain
Cue 61 — 00:03:53,800 --> 00:03:55,960
attacks if you operate this proxy.
Cue 62 — 00:03:56,000 --> 00:04:00,240
Read also the blog, not only the release notes, because the blog contains a lot of
Cue 63 — 00:04:00,280 --> 00:04:03,460
lessons learned in production and some fixes that you can apply.
Cue 64 — 00:04:03,520 --> 00:04:09,000
One of the CLI params that we are adding is we are restarting the LiteLLM pods
Cue 65 — 00:04:09,040 --> 00:04:12,920
basically after 20,000 requests, in order to combat any memory leaks that exist
Cue 66 — 00:04:12,960 --> 00:04:18,280
there. We hope that the rewrite in Rust will help a lot with this stability
Cue 67 — 00:04:18,320 --> 00:04:23,480
problems. We built also a lot of tooling, and our community has been very active
Cue 68 — 00:04:23,520 --> 00:04:27,120
with that. And some of those tools were early stage.
Cue 69 — 00:04:27,160 --> 00:04:29,320
When coding agents did not exist.
Cue 70 — 00:04:29,360 --> 00:04:35,160
We built a CLI tool that provides users in early, early and easy way of sending a
Cue 71 — 00:04:35,200 --> 00:04:41,000
prompt to the models and using Unix pipes in order to pipe files there and have
Cue 72 — 00:04:41,040 --> 00:04:46,100
first experiences on how basically models will reply to those changes.
Cue 73 — 00:04:46,480 --> 00:04:52,360
We also invested a lot in the security bit, so if you read, tutorials on the
Cue 74 — 00:04:52,400 --> 00:04:53,680
internet, they often will tell you.
Cue 75 — 00:04:53,720 --> 00:04:58,140
Oh, you can paste the token hardcoded in the authorization header of the MCP.
Cue 76 — 00:04:58,240 --> 00:05:04,000
When we started hosting MCP servers, we explicitly added a functionality of a MCP
Cue 77 — 00:05:04,040 --> 00:05:09,440
bridge into the CLI tool that basically allows to convert a remote server to a local
Cue 78 — 00:05:09,480 --> 00:05:12,960
server with standard input output, transport.
Cue 79 — 00:05:13,000 --> 00:05:16,840
And this allowed users to basically use this in VS code without the need to care
Cue 80 — 00:05:16,880 --> 00:05:20,440
about tokens, because for trusted hosts they are auto injected.
Cue 81 — 00:05:20,480 --> 00:05:22,360
We also use it to configure coding agents.
Cue 82 — 00:05:22,400 --> 00:05:26,600
We have configurations that basically you can apply their version versioned in git,
Cue 83 — 00:05:26,640 --> 00:05:31,140
and you can use one of the commands of the CLI to apply them to your local machine.
Cue 84 — 00:05:31,200 --> 00:05:35,440
Our community has also been very active in writing plugins for OpenCode and Pi,
Cue 85 — 00:05:35,480 --> 00:05:38,360
and that basically are the model providers on LiteLLM.
Cue 86 — 00:05:38,400 --> 00:05:42,600
They also fix some compatibility issues between the harness and the model
Cue 87 — 00:05:42,640 --> 00:05:47,380
information that is provided, for example, for reasoning effort parameter mapping.
Cue 88 — 00:05:47,640 --> 00:05:52,200
What we also discovered more recently with OpenCode version 2 is that it does weird
Cue 89 — 00:05:52,240 --> 00:05:56,000
things. When you wake up your laptop, it sends parallel requests to the model
Cue 90 — 00:05:56,040 --> 00:05:59,520
endpoints. It sends those requests when the networking is not ready.
Cue 91 — 00:05:59,560 --> 00:06:04,800
So in our providers, we had to add file based fallback caches in order to combat the
Cue 92 — 00:06:04,840 --> 00:06:08,800
problem and to not have to plug and fail, because then you need to restart the
Cue 93 — 00:06:08,840 --> 00:06:12,660
service on OpenCode. And it's very difficult to provide support for users.
Cue 94 — 00:06:13,520 --> 00:06:17,960
What we also discovered early on is that, there are users who are building custom
Cue 95 — 00:06:18,000 --> 00:06:22,320
agents against the proxy with Python AI, and they do not know that prompt caching
Cue 96 — 00:06:22,360 --> 00:06:24,920
exists. So we added prompt caching on the proxy.
Cue 97 — 00:06:24,960 --> 00:06:29,080
But we also figured that it would be great to have a tool that would help users
Cue 98 — 00:06:29,120 --> 00:06:31,080
understand that they are doing something wrong.
Cue 99 — 00:06:31,120 --> 00:06:33,680
So we wrote another proxy that you run locally.
Cue 100 — 00:06:33,720 --> 00:06:38,520
This proxy is providing you with a simple TUI that is calculating the token costs
Cue 101 — 00:06:38,560 --> 00:06:43,880
here shown on the picture, and it's marking with the zeros with a red sign that
Cue 102 — 00:06:43,920 --> 00:06:47,240
basically show users, hey, you're not using prompt caching at all,
Cue 103 — 00:06:47,280 --> 00:06:49,800
so maybe there is something wrong in the way you're using it,
Cue 104 — 00:06:49,840 --> 00:06:53,580
and you have to rewrite your agent to inject the prompt caching headers.
Cue 105 — 00:06:53,800 --> 00:06:58,260
We also have request response logging in the proxy that helps when debugging.
Cue 106 — 00:06:58,800 --> 00:07:02,000
problems with LiteLLM because you can take the request responses,
Cue 107 — 00:07:02,040 --> 00:07:06,320
your session logs the LiteLLM codebase and ask Claude to figure out where the bug
Cue 108 — 00:07:06,360 --> 00:07:12,000
is. We also have an authorization header, injection, and proxy that basically helps
Cue 109 — 00:07:12,040 --> 00:07:17,320
users to use any other coding agent that is not supporting the API key helper
Cue 110 — 00:07:17,360 --> 00:07:20,920
pattern, right? We have access to the proxy gated by OAuth tokens.
Cue 111 — 00:07:20,960 --> 00:07:25,820
They expire. So it's a little bit tricky to always export a new,
Cue 112 — 00:07:26,000 --> 00:07:28,400
environment variable and restart the coding harness.
Cue 113 — 00:07:28,440 --> 00:07:30,400
Nobody wants to do that. Right.
Cue 114 — 00:07:30,440 --> 00:07:32,280
So now coming to the impact right.
Cue 115 — 00:07:32,320 --> 00:07:35,840
Our users have been getting access to various different models through a proxy.
Cue 116 — 00:07:35,880 --> 00:07:37,600
They can use all those different tools.
Cue 117 — 00:07:37,640 --> 00:07:41,720
What can they do. Well we've seen leverage and platform teams.
Cue 118 — 00:07:41,760 --> 00:07:46,360
We have a fleet mindset where we basically have a platform team that can execute
Cue 119 — 00:07:46,400 --> 00:07:52,760
changes all across our code bases, which then results in thousands of PRs issued
Cue 120 — 00:07:52,800 --> 00:07:58,043
across the code base and LLMs have provided the teams that are issuing those PRs
Cue 121 — 00:07:58,083 --> 00:08:00,480
with transformations that didn't exist before, right?
Cue 122 — 00:08:00,520 --> 00:08:02,080
So they can cover more ground.
Cue 123 — 00:08:02,120 --> 00:08:06,040
And in addition to the deterministic transformations that they've been doing before,
Cue 124 — 00:08:06,080 --> 00:08:09,920
they can now do LLM-based transformations that account more for the variability of
Cue 125 — 00:08:09,960 --> 00:08:13,880
the changes that the engineering teams are doing in their code bases and help with
Cue 126 — 00:08:13,920 --> 00:08:18,760
platform feature adoption. We also have seen great success with migrations.
Cue 127 — 00:08:18,800 --> 00:08:23,280
We do migrations in a way that a team that is imposing the migration on all other
Cue 128 — 00:08:23,320 --> 00:08:26,960
teams has to do the migration themselves, right?
Cue 129 — 00:08:27,000 --> 00:08:29,840
So they basically write skills, they write helper functions,
Cue 130 — 00:08:29,880 --> 00:08:33,700
and they issue using the mechanism that I mentioned in the slide before.
Cue 131 — 00:08:34,000 --> 00:08:36,160
PRs that the teams have to accept.
Cue 132 — 00:08:36,200 --> 00:08:40,720
Right. So there is no more migration that you basically send an email to your team
Cue 133 — 00:08:40,760 --> 00:08:45,640
and say, oh my dear client, please update this as a new endpoint,
Cue 134 — 00:08:45,680 --> 00:08:47,640
right? You have to execute the migration yourself.
Cue 135 — 00:08:47,680 --> 00:08:53,680
And what we have seen is on large migrations, for example, 200 database tables that
Cue 136 — 00:08:53,720 --> 00:08:56,640
needed to be migrated with 50 data pipelines.
Cue 137 — 00:08:56,680 --> 00:08:59,740
The effort on those migrations was reduced almost by half.
Cue 138 — 00:09:00,680 --> 00:09:04,080
Now all this code that is being created, including those automated PRs,
Cue 139 — 00:09:04,120 --> 00:09:08,200
of course, has some number of PRs that are being created.
Cue 140 — 00:09:08,240 --> 00:09:11,200
But if you look deeper, there is another picture emerging.
Cue 141 — 00:09:11,240 --> 00:09:15,200
What you can see on this graph is the share of PRs in terms of the size
Cue 142 — 00:09:15,240 --> 00:09:19,640
distribution. Right. So on the y axis you have the percentage on the x axis you have
Cue 143 — 00:09:19,680 --> 00:09:26,480
different size buckets. And what we see in quarterly numbers is that since 2025
Cue 144 — 00:09:26,520 --> 00:09:32,060
quarter three, we have a trend where the large PRs are heavily increasing in size,
Cue 145 — 00:09:32,120 --> 00:09:35,960
even though we have told our teams, hey, there is a PR size that you probably should
Cue 146 — 00:09:36,000 --> 00:09:41,240
agree, from which point in time you should be splitting your PRs and not imposing
Cue 147 — 00:09:41,280 --> 00:09:43,200
thousands of lines of codes on you.
Cue 148 — 00:09:43,240 --> 00:09:47,440
This doesn't seem to be a trend that we can stop this way, right?
Cue 149 — 00:09:47,480 --> 00:09:52,780
And to deal with those PRs, we have a risk based PR scoring approach.
Cue 150 — 00:09:52,840 --> 00:09:57,360
In this approach, what we are looking at is what is the likelihood that the PR,
Cue 151 — 00:09:57,400 --> 00:10:01,220
if it's going to be merged and rolled out, will cause a problem in production,
Cue 152 — 00:10:01,680 --> 00:10:07,740
right. And at least half, one third of the PR the teams are having are low risk.
Cue 153 — 00:10:07,800 --> 00:10:12,640
What does low risk mean for us basically means that, you get an automatic approval
Cue 154 — 00:10:12,680 --> 00:10:15,860
from the bot. And as an author you can merge this PR.
Cue 155 — 00:10:16,120 --> 00:10:19,400
If you do merge this PR and you adapt to this new way of working,
Cue 156 — 00:10:19,440 --> 00:10:23,080
there is a 20 to 40% lead time reduction on those.
Cue 157 — 00:10:23,120 --> 00:10:25,440
And our engineers love this bot so much.
Cue 158 — 00:10:25,480 --> 00:10:28,480
Many of them say it's a game changer on how they work, because they don't have to
Cue 159 — 00:10:28,520 --> 00:10:30,340
beg for the rubber stamps anymore.
Cue 160 — 00:10:30,480 --> 00:10:34,920
And it's also changing behavior of people because they see medium risk and then they
Cue 161 — 00:10:34,960 --> 00:10:39,000
try to think, okay, what do I need to do so that this PR becomes low risk and they
Cue 162 — 00:10:39,040 --> 00:10:43,080
are splitting changes into smaller pieces so that they can move ahead with the low
Cue 163 — 00:10:43,120 --> 00:10:45,200
risk change that will be an additive change.
Cue 164 — 00:10:45,240 --> 00:10:48,640
And for breaking changes, they need the approval and discussion with the team how
Cue 165 — 00:10:48,680 --> 00:10:50,420
this is going to be rolled out.
Cue 166 — 00:10:50,680 --> 00:10:51,960
How does it look for the user?
Cue 167 — 00:10:52,000 --> 00:10:55,700
You basically get a comment from the bot triggered by a git webhook,
Cue 168 — 00:10:56,320 --> 00:11:00,980
and if it's going to be high risk, you get the information that sorry,
Cue 169 — 00:11:01,120 --> 00:11:05,640
no approval for you. What are some of the rules that we have encoded?
Cue 170 — 00:11:05,680 --> 00:11:10,120
We basically scanned our postmortems, try to understand what infrastructure changes
Cue 171 — 00:11:10,160 --> 00:11:15,840
are, high risk, and what infrastructure changes typically lead to production issues.
Cue 172 — 00:11:15,880 --> 00:11:20,240
Right. And it's complicated rollouts, multi-stage rollouts, not using feature flags
Cue 173 — 00:11:20,280 --> 00:11:24,120
because you then, cannot rollback changes quickly.
Cue 174 — 00:11:24,160 --> 00:11:28,800
Or in this case, a database change where you add an index without concurrently work.
Cue 175 — 00:11:28,840 --> 00:11:32,960
Right. And if you are going to fix some of the messages that are being that are
Cue 176 — 00:11:33,000 --> 00:11:37,280
being raised by the higher risk PRs, you have a chance that you will go down to
Cue 177 — 00:11:37,320 --> 00:11:41,160
medium risk. If you split, maybe you will go even to low risk and then you can ship
Cue 178 — 00:11:41,200 --> 00:11:45,320
faster, right? So that's one example of the guardrails that we have.
Cue 179 — 00:11:45,360 --> 00:11:49,540
But there is a little bit more question is why do we need to have guardrails.
Cue 180 — 00:11:49,640 --> 00:11:51,680
That's the cost distribution right.
Cue 181 — 00:11:51,720 --> 00:11:54,140
On the x axis you have the different users.
Cue 182 — 00:11:54,720 --> 00:11:57,840
On the y axis you have logarithmic cost right.
Cue 183 — 00:11:57,880 --> 00:12:00,680
It basically means that it's growing very quickly.
Cue 184 — 00:12:00,720 --> 00:12:05,880
If you look at the long tail of users and it's something that we observe,
Cue 185 — 00:12:05,920 --> 00:12:11,360
right. We look at the P50; the relation to P90 is 10x.
Cue 186 — 00:12:11,400 --> 00:12:15,200
Right. So the high spenders are spending really a lot.
Cue 187 — 00:12:15,240 --> 00:12:18,040
And I get the privilege to talk to them right.
Cue 188 — 00:12:18,080 --> 00:12:22,920
And I can learn a lot of the patterns that users are applying just by talking to the
Cue 189 — 00:12:22,960 --> 00:12:26,760
top spenders and trying to understand, like what they have been trying to do.
Cue 190 — 00:12:26,800 --> 00:12:28,860
In many of the cases, it's like
Cue 191 — 00:12:30,200 --> 00:12:32,560
users didn't know that it costs so much, right?
Cue 192 — 00:12:32,600 --> 00:12:34,440
So they didn't have visibility on the cost.
Cue 193 — 00:12:34,480 --> 00:12:38,240
They didn't have an understanding on what they can do to reduce those costs.
Cue 194 — 00:12:38,280 --> 00:12:42,760
And my objective is that the same person doesn't meet me in the next month.
Cue 195 — 00:12:42,800 --> 00:12:47,480
And trying to trying to discuss again what is their use case that they are having.
Cue 196 — 00:12:47,520 --> 00:12:50,560
Right. And we have a safety cap that we apply with the safety cap.
Cue 197 — 00:12:50,600 --> 00:12:56,240
They basically get the experience of, sorry, the API is out of budget for you,
Cue 198 — 00:12:56,280 --> 00:13:00,240
so please contact us. Right. And the user can choose to they want this or do they
Cue 199 — 00:13:00,280 --> 00:13:01,600
say okay, maybe I used it too much.
Cue 200 — 00:13:01,640 --> 00:13:04,340
I don't want to. I have more of the juice.
Cue 201 — 00:13:04,640 --> 00:13:09,560
And we use two tools Agents View for viewing the sessions and Coburn for viewing the
Cue 202 — 00:13:09,600 --> 00:13:14,960
local costs in order to give users tools that allow them to understand what to do.
Cue 203 — 00:13:15,000 --> 00:13:20,280
And we have great success in moving away from Anthropic models and guiding users to
Cue 204 — 00:13:20,320 --> 00:13:24,880
use GPT models trying out harnesses like OpenCode, and they typically then share the
Cue 205 — 00:13:24,920 --> 00:13:30,440
lessons. Hey, I reduced my bill so much that I can show you how right and cost per
Cue 206 — 00:13:30,480 --> 00:13:32,040
user is something that we want to protect.
Cue 207 — 00:13:32,080 --> 00:13:36,840
This is what we budget against, and we protect quality by looking at the lead time
Cue 208 — 00:13:36,880 --> 00:13:41,040
that we see decreasing and the change failure rate that we see throughout the year.
Cue 209 — 00:13:41,080 --> 00:13:46,560
Being in the stable corridor, the experience of being blocked on an API is of course
Cue 210 — 00:13:46,600 --> 00:13:50,680
not a great one. So we have proactive notifications that we send to users that warn
Cue 211 — 00:13:50,720 --> 00:13:53,760
them that something is coming their way, unless they're going to change their
Cue 212 — 00:13:53,800 --> 00:13:58,240
patterns. And we link to documentation in the emails, that helps them understand how
Cue 213 — 00:13:58,280 --> 00:13:59,555
to deal with that.
Cue 225 — 00:14:41,214 --> 00:14:44,328
We also have a lot of tools that our engineers are trying out,
Cue 226 — 00:14:44,368 --> 00:14:46,088
and we apply the tech radar concept there.
Cue 227 — 00:14:46,128 --> 00:14:51,048
We have an AI section in the tech radar, where we are basically tracking all the
Cue 228 — 00:14:51,088 --> 00:14:56,648
tools that we are trying out and where we are sharing, which of those those tools
Cue 229 — 00:14:56,688 --> 00:15:01,568
have been tried out more because they are in trial, and maybe they are already ready
Cue 230 — 00:15:01,608 --> 00:15:06,368
for some of the wider adoption, because we have written plugins for those tools,
Cue 231 — 00:15:06,408 --> 00:15:10,168
because we have written model providers, and we can recommend those tools to be
Cue 232 — 00:15:10,208 --> 00:15:13,888
used. It's also a mechanism that we use to tell to our engineers that,
Cue 233 — 00:15:13,928 --> 00:15:17,768
hey, some of those tools are like really early stage and they have not proven its
Cue 234 — 00:15:17,808 --> 00:15:21,888
value. So even if you read headlines on social media, it's not something that you
Cue 235 — 00:15:21,928 --> 00:15:26,248
have to use immediately. You can wait for the trial and you can read on what is
Cue 236 — 00:15:26,288 --> 00:15:28,028
happening with the Assess entries.
Cue 237 — 00:15:28,168 --> 00:15:31,268
We also have an open source license licensing guide.
Cue 238 — 00:15:31,888 --> 00:15:34,608
There are lots of models that are coming out with modified MIT,
Cue 239 — 00:15:34,648 --> 00:15:37,368
and you have to understand when can you use K3?
Cue 240 — 00:15:37,408 --> 00:15:39,328
Can you use it for customer facing use cases?
Cue 241 — 00:15:39,368 --> 00:15:41,108
Can you use it to work with code?
Cue 242 — 00:15:42,088 --> 00:15:46,648
We support our teams that are building AI use cases with some legal self-assessment
Cue 243 — 00:15:46,688 --> 00:15:51,328
tools, so they have a form when they fill out the form, they have an assessment like
Cue 244 — 00:15:51,368 --> 00:15:53,768
is your use case low risk? Can you proceed?
Cue 245 — 00:15:53,808 --> 00:15:56,908
Or maybe do you need to have a legal review on top of that?
Cue 246 — 00:15:57,128 --> 00:16:02,168
And basically some requirements that we have in Europe for the AI Act also require
Cue 247 — 00:16:02,208 --> 00:16:06,688
us to track model usage and the AI systems and the use cases that that they are
Cue 248 — 00:16:06,728 --> 00:16:11,088
connected with. We have a developer portal based on Backstage where we are
Cue 249 — 00:16:11,128 --> 00:16:15,768
collecting those models, and on top of that, we also verify whether models that go
Cue 250 — 00:16:15,808 --> 00:16:21,128
into production through a Docker scan, have also been registered in the developer
Cue 251 — 00:16:21,168 --> 00:16:26,488
portal earlier. For our customer facing use cases, we have offline and online
Cue 252 — 00:16:26,528 --> 00:16:30,268
evaluations that allow us to more easily switch models.
Cue 253 — 00:16:30,328 --> 00:16:35,888
And for the use cases that we have running in our shop for the B2C,
Cue 254 — 00:16:35,928 --> 00:16:40,448
we have probes that run in production and constantly probe for common use cases like
Cue 255 — 00:16:40,488 --> 00:16:43,968
adding to cart or viewing the catalog view in the product detail page,
Cue 256 — 00:16:44,008 --> 00:16:48,448
and they page the on-call teams when there is instability in the system.
Cue 257 — 00:16:48,488 --> 00:16:52,168
Right. That's a mechanism that we had already running in the past,
Cue 258 — 00:16:52,208 --> 00:16:55,608
but it proved its value and of the investment that we have done.
Cue 259 — 00:16:55,648 --> 00:16:59,792
And we want to expand this mechanism to also be triggered as part of the CD
Cue 260 — 00:16:59,832 --> 00:17:04,728
pipelines for, the UI's that we are developing for our customers in the Zalando app
Cue 261 — 00:17:04,768 --> 00:17:07,568
and website and also for the backend services.
Cue 262 — 00:17:07,608 --> 00:17:12,408
But for the backend services, it's a little bit more advanced to connect the rollout
Cue 263 — 00:17:12,448 --> 00:17:14,008
that you're doing to the actual probe.
Cue 264 — 00:17:14,048 --> 00:17:17,648
So we want to annotate our backend services with the critical business operations
Cue 265 — 00:17:17,688 --> 00:17:22,308
that they support, and then be able to isolate test traffic so that we can run,
Cue 266 — 00:17:22,488 --> 00:17:26,728
basically end-to-end probes from the customer facing application and then route the
Cue 267 — 00:17:26,768 --> 00:17:30,428
traffic to the backend that has a test deployment when it's needed.
Cue 268 — 00:17:31,808 --> 00:17:36,068
Now coming to further scaling, who's building an agent platform here?
Cue 269 — 00:17:36,448 --> 00:17:39,168
Right. We are building one. And the question is why.
Cue 270 — 00:17:39,208 --> 00:17:42,008
Right. First thing is we want to have hosted agents.
Cue 271 — 00:17:42,048 --> 00:17:47,008
And ideally we can also have those agents in the slack on Kubernetes that we have
Cue 272 — 00:17:47,048 --> 00:17:50,068
and therefore be cost neutral to our operations.
Cue 273 — 00:17:50,168 --> 00:17:55,288
We also have external use cases where for our partners, we are offering MCP servers,
Cue 274 — 00:17:55,328 --> 00:18:00,288
and they need a little bit of infrastructure to be exposed in a secure way and
Cue 275 — 00:18:00,328 --> 00:18:02,508
integrated with our IAM systems.
Cue 276 — 00:18:02,608 --> 00:18:09,568
We also have, through the platform, a registry of the MCP servers and the skill
Cue 277 — 00:18:09,608 --> 00:18:11,448
marketplaces that our teams are having.
Cue 278 — 00:18:11,488 --> 00:18:14,328
So basically, if you have a skill marketplace on team level,
Cue 279 — 00:18:14,368 --> 00:18:21,368
on, our company, content, or company skills that we created
Cue 280 — 00:18:21,408 --> 00:18:25,248
or maybe on department level, you basically can install a GitHub app.
Cue 281 — 00:18:25,288 --> 00:18:29,168
The GitHub app would slurp those skills, and then they will be available to use on
Cue 282 — 00:18:29,208 --> 00:18:34,228
the platform. For building the platform, we composed open source components,
Cue 283 — 00:18:34,608 --> 00:18:38,568
kagent for the runtime, agentgateway for model access and routing.
Cue 284 — 00:18:38,608 --> 00:18:43,408
And we had to solve the identity problem on our own by building an identity broker
Cue 285 — 00:18:43,448 --> 00:18:49,248
that I will introduce next. For some add ons, we use Cowra as a Postgres based
Cue 286 — 00:18:49,288 --> 00:18:54,168
memory system. We had better experience with this than with Mem0,
Cue 287 — 00:18:54,208 --> 00:18:58,848
and our agents that are built on the platform can be shared on team level with
Cue 288 — 00:18:58,888 --> 00:19:02,428
individuals, or be public to all users of the platform itself.
Cue 289 — 00:19:02,608 --> 00:19:08,128
And what we also want to do to ease the adoption of the platform with our users is
Cue 290 — 00:19:08,168 --> 00:19:13,868
to have, templates for common use cases that, teams can start off as a baseline.
Cue 291 — 00:19:14,328 --> 00:19:20,488
Now, the identity broker. So what is it basically an identity infrastructure for
Cue 292 — 00:19:20,528 --> 00:19:23,288
agents that can act on behalf of a user.
Cue 293 — 00:19:23,328 --> 00:19:29,028
Right. It contains a token vault, and the tokens are exchanged transparently
Cue 294 — 00:19:29,328 --> 00:19:31,648
whenever they are needed by the runtime.
Cue 295 — 00:19:31,688 --> 00:19:36,128
When it's executing tool calls against the MCP server and the user who is using the
Cue 296 — 00:19:36,168 --> 00:19:39,208
agent is only needing to have the user token.
Cue 297 — 00:19:39,248 --> 00:19:41,408
So how does the system work under the hood?
Cue 298 — 00:19:41,448 --> 00:19:46,608
Basically, an an agent that is registered on the platform defines what permissions
Cue 299 — 00:19:46,648 --> 00:19:51,568
it requires in order to function, for example, to access internal APIs or to be able
Cue 300 — 00:19:51,608 --> 00:19:55,568
to access third party systems like Google, GitHub or Databricks.
Cue 301 — 00:19:55,608 --> 00:20:00,028
And then what happens is, whenever the user is using the agent for the first time,
Cue 302 — 00:20:00,408 --> 00:20:03,528
they are going to be presented with a consent screen.
Cue 303 — 00:20:03,568 --> 00:20:07,048
Then on the consent screen, they will log in to all those third party systems that I
Cue 304 — 00:20:07,088 --> 00:20:11,928
mentioned. And every time they complete the login, the token will be stored in the
Cue 305 — 00:20:11,968 --> 00:20:17,488
vault and will be available for use when the user finally approves the delegation.
Cue 306 — 00:20:17,528 --> 00:20:22,088
And then every time there is a tool call being made, there is a transparent exchange
Cue 307 — 00:20:22,128 --> 00:20:26,388
that happens on the edge, the which makes the whole system more secure.
Cue 308 — 00:20:26,528 --> 00:20:30,328
The system is open source, released it like a month or two ago.
Cue 309 — 00:20:30,368 --> 00:20:34,537
It's written spec-driven, so you can use the specs in order to rewrite it in Rust if
Cue 310 — 00:20:34,577 --> 00:20:40,448
you don't like Golang. From the user view, the agent platform has a UI that makes it
Cue 311 — 00:20:40,488 --> 00:20:42,848
simple for non-technical users to use it as well.
Cue 312 — 00:20:42,888 --> 00:20:47,408
So all you have, all you can do, in the UI is basically you define what is your
Cue 313 — 00:20:47,448 --> 00:20:49,928
agent name, who is owning it? Are you sharing it?
Cue 314 — 00:20:49,968 --> 00:20:53,008
You define the prompt. You say, what is the memory level that you use?
Cue 315 — 00:20:53,048 --> 00:20:58,038
Non private for a user scoped memory or team?
Cue 316 — 00:20:58,568 --> 00:21:02,208
Shared for being shared across all users of the agent.
Cue 317 — 00:21:02,248 --> 00:21:05,288
And then what you can do is you can pick from the MCP servers that are registered on
Cue 318 — 00:21:05,328 --> 00:21:08,968
the platform. That's the simplest manifestation of an agent that we have on the
Cue 319 — 00:21:09,008 --> 00:21:14,916
platform. And what happens then when you save the agent is we have KRO as a resource
Cue 320 — 00:21:14,956 --> 00:21:18,448
orchestrator that will, on Kubernetes, create the necessary resources for this agent
Cue 321 — 00:21:18,488 --> 00:21:22,608
to function. Right. So it will contain a config map with the configuration,
Cue 322 — 00:21:22,648 --> 00:21:25,268
the agent resources for the agent and it's up.
Cue 323 — 00:21:25,648 --> 00:21:30,348
It will orchestrate the identity broker by creating the necessary resources and
Cue 324 — 00:21:30,648 --> 00:21:36,568
linking up the delegations. And then it will set up an A2A route and A2A policy so
Cue 325 — 00:21:36,608 --> 00:21:39,388
that the agents can be actually called over the APIs.
Cue 326 — 00:21:39,688 --> 00:21:44,308
Right. And there is also another reason why we have built the platform is,
Cue 327 — 00:21:44,408 --> 00:21:47,848
we have a lot of users who use chat, but we use Google Chat,
Cue 328 — 00:21:47,888 --> 00:21:50,528
we don't get all the cool bots and nobody is writing plugins,
Cue 329 — 00:21:50,568 --> 00:21:52,248
so we have to write it on our own.
Cue 330 — 00:21:52,288 --> 00:21:56,608
And we basically have a, bot that you can configure in each of the chat rooms and
Cue 331 — 00:21:56,648 --> 00:21:59,768
you just say, what is the agent name that is going to be active there?
Cue 332 — 00:21:59,808 --> 00:22:03,528
We launched the platform two weeks ago, and we have more than 300 agents that people
Cue 333 — 00:22:03,568 --> 00:22:07,848
have created. And you can see the full stats and the KPI dashboard.
Cue 334 — 00:22:07,888 --> 00:22:10,888
And the questions that we are asking ourselves now is like, also,
Cue 335 — 00:22:10,928 --> 00:22:12,768
how many of those are duplicates?
Cue 336 — 00:22:12,808 --> 00:22:16,592
How do we need to deal with like reviewing for agents because it bypasses the
Cue 337 — 00:22:16,632 --> 00:22:18,768
four-eyes principle for now, because it can be private.
Cue 338 — 00:22:18,808 --> 00:22:22,328
But then maybe you share and as part of sharing you need to have an approval.
Cue 339 — 00:22:22,368 --> 00:22:25,728
And maybe we like the same way as we do the risk scoring on PRs.
Cue 340 — 00:22:25,768 --> 00:22:29,568
We also have to give feedback to users, whether the agent that they are creating
Cue 341 — 00:22:29,608 --> 00:22:33,088
with the current capabilities that are available on the platform has a chance to
Cue 342 — 00:22:33,128 --> 00:22:37,728
work. Now coming to enabling our non-technical users, right?
Cue 343 — 00:22:37,768 --> 00:22:39,968
Our citizen developers, those are great people.
Cue 344 — 00:22:40,008 --> 00:22:44,148
They have many ideas and many of the times they are gated by tooling access,
Cue 345 — 00:22:44,208 --> 00:22:47,168
right. So they were also never onboarded developers.
Cue 346 — 00:22:47,208 --> 00:22:51,468
So they don't know that deployment means the same as publishing.
Cue 347 — 00:22:51,528 --> 00:22:57,048
And they end up making AI assisted PR, sometimes willingly, willingly,
Cue 348 — 00:22:57,088 --> 00:23:00,608
and sometimes because there is a skill that is helping them to make them.
Cue 349 — 00:23:00,648 --> 00:23:05,368
For example, in our knowledge bases wiki style, we have skills that basically end up
Cue 350 — 00:23:05,408 --> 00:23:08,588
being fired whenever the user says you are wrong.
Cue 351 — 00:23:08,888 --> 00:23:13,168
And then, citizen developers also tend to solve everything with the agent.
Cue 352 — 00:23:13,208 --> 00:23:15,908
So if they encounter any friction of the platform.
Cue 353 — 00:23:16,248 --> 00:23:19,688
Then they will use a route that the agent recommends, not that the platform team
Cue 354 — 00:23:19,728 --> 00:23:22,888
recommends. So they need help and productized tools.
Cue 355 — 00:23:22,928 --> 00:23:29,248
One example tool that our design team has built, is a playground for prototypes.
Cue 356 — 00:23:29,288 --> 00:23:31,248
So they have taken the design system.
Cue 357 — 00:23:31,288 --> 00:23:35,368
They have taken and created templates for user experiences.
Cue 358 — 00:23:35,408 --> 00:23:39,868
And then what you can do is you can prompt prompt a prototype into existence.
Cue 359 — 00:23:39,928 --> 00:23:42,248
Right. And the way you do this is you go to a repository.
Cue 360 — 00:23:42,288 --> 00:23:45,528
This repository has a skill that is called prototype.
Cue 361 — 00:23:45,568 --> 00:23:48,808
And you basically will be able to have this prototype created.
Cue 362 — 00:23:48,848 --> 00:23:51,048
There are a few problems that are listed here.
Cue 363 — 00:23:51,088 --> 00:23:53,648
Probably users that are non-technical.
Cue 364 — 00:23:53,688 --> 00:23:56,328
They don't want to check out a repo and have Claude on their machines,
Cue 365 — 00:23:56,368 --> 00:24:00,608
so maybe they just need a agent on our agent platform and set with that.
Cue 366 — 00:24:00,648 --> 00:24:05,808
Then there is a mix between the tool that was created and the prototype definitions.
Cue 367 — 00:24:05,848 --> 00:24:08,308
So you would want to separate this as well.
Cue 368 — 00:24:08,448 --> 00:24:12,128
And the tooling is internal only gated by our IAM system.
Cue 369 — 00:24:12,168 --> 00:24:16,568
But the design team actually wants to publish those so that they have user research
Cue 370 — 00:24:16,608 --> 00:24:20,148
opportunities with external users who will not have an account with us.
Cue 371 — 00:24:20,248 --> 00:24:23,168
Right. So we are working with the team to productize this use case,
Cue 372 — 00:24:23,208 --> 00:24:27,048
and we have more and more use cases where people want to publish a prototype.
Cue 373 — 00:24:27,088 --> 00:24:30,128
So far would they have been doing is they have been going the developer route,
Cue 374 — 00:24:30,168 --> 00:24:34,568
creating a repository using our static site hosting for that and having prototypes
Cue 375 — 00:24:34,608 --> 00:24:37,928
in the browser. But it's not a great route because you have to start with the
Cue 376 — 00:24:37,968 --> 00:24:43,148
repository. And who will know as a non developer that a repository needs to exist.
Cue 377 — 00:24:44,248 --> 00:24:45,668
Now coming to enablement.
Cue 378 — 00:24:47,448 --> 00:24:52,888
We have some challenges that our developers are telling us that basically the pace
Cue 379 — 00:24:52,928 --> 00:24:54,908
is way too high for them to keep up.
Cue 380 — 00:24:55,008 --> 00:24:59,088
They just learned a new tool. And then there is a new technique and new kid on the
Cue 381 — 00:24:59,128 --> 00:25:01,968
block, and everybody is saying, oh, you need to have this approach.
Cue 382 — 00:25:02,008 --> 00:25:04,248
And they also say that they have insufficient time to learn,
Cue 383 — 00:25:04,288 --> 00:25:08,308
because the pressure on delivery is so high that they cannot take a step back.
Cue 384 — 00:25:08,568 --> 00:25:12,648
And when we do unconference and exchanges, what they say is,
Cue 385 — 00:25:12,688 --> 00:25:16,808
oh, whenever I change with my peers, I kind of feel relief because I know I'm not
Cue 386 — 00:25:16,848 --> 00:25:18,748
the only one who is struggling.
Cue 387 — 00:25:19,048 --> 00:25:23,928
And what they also say is, what is our job looking like in the next 2 or 3 years?
Cue 388 — 00:25:23,968 --> 00:25:27,408
They're asking for guidance, but we are not really in a place where we can provide
Cue 389 — 00:25:27,448 --> 00:25:32,308
this guidance. So our enablement so far has been we have our,
Cue 390 — 00:25:32,728 --> 00:25:34,888
guild sessions, knowledge sharing sessions.
Cue 391 — 00:25:34,928 --> 00:25:38,488
We have more than 120 people who are attending them on a weekly basis.
Cue 392 — 00:25:38,528 --> 00:25:42,568
And this is like for the AI guild people, for the people who need to know what is
Cue 393 — 00:25:42,608 --> 00:25:46,088
the newest things, what have we released in terms of platform capabilities,
Cue 394 — 00:25:46,128 --> 00:25:49,688
what has been available on the market, and people exchange there and they are
Cue 395 — 00:25:49,728 --> 00:25:54,408
basically super engaged. What we then seek is from the people of presenting there
Cue 396 — 00:25:54,448 --> 00:25:57,648
who can help us devise and create training formats.
Cue 397 — 00:25:57,688 --> 00:26:00,408
So we have a format that we call labs and the labs.
Cue 398 — 00:26:00,448 --> 00:26:05,048
We are trying things out in the exploratory phase or in the phase where we have a
Cue 399 — 00:26:05,088 --> 00:26:09,968
rough idea about the tutorial, where we pair up people in pairs and they try to go
Cue 400 — 00:26:10,008 --> 00:26:12,848
through this tutorial and they give us feedback if this is something that helped
Cue 401 — 00:26:12,888 --> 00:26:17,568
them learn something new. If it did, we promote it to a training that we then run
Cue 402 — 00:26:17,608 --> 00:26:22,808
monthly. One example of such trainings is how to build agents with AI that basically
Cue 403 — 00:26:22,848 --> 00:26:27,368
give developers and information and the necessary knowledge foundation to understand
Cue 404 — 00:26:27,408 --> 00:26:31,028
how the coding agents are working under the hood and how loops work as well.
Cue 405 — 00:26:31,368 --> 00:26:34,568
We are now in Q4, rolling out a pilot for our engineers.
Cue 406 — 00:26:34,608 --> 00:26:37,368
That will be a four week training program.
Cue 407 — 00:26:37,408 --> 00:26:40,108
This training program will start with a kickoff on Monday.
Cue 408 — 00:26:40,448 --> 00:26:44,448
And then, with the knowledge that we have provided during this kickoff,
Cue 409 — 00:26:44,488 --> 00:26:48,457
people will work on items from their backlog trying to apply the techniques that we
Cue 410 — 00:26:48,497 --> 00:26:52,368
saw in a time-boxed manner, 1 or 2 hours, and they will work in pairs.
Cue 411 — 00:26:52,408 --> 00:26:56,848
The engineering manager can choose, do they pick a person that is a skeptic and give
Cue 412 — 00:26:56,888 --> 00:26:59,048
them pair them with an expert?
Cue 413 — 00:26:59,088 --> 00:27:03,008
Or do they pair people in a different way, but they own the training.
Cue 414 — 00:27:03,048 --> 00:27:04,768
They own the assignment of the issues.
Cue 415 — 00:27:04,808 --> 00:27:09,468
And with this, we also delegate a little bit of the ownership to the EMs.
Cue 416 — 00:27:09,768 --> 00:27:14,488
And we also support our teams through this program with a readiness scan for
Cue 417 — 00:27:14,528 --> 00:27:19,168
repositories. Right. So we have some patterns and platform features that we
Cue 418 — 00:27:19,208 --> 00:27:23,688
developed over time, like gradual deployments that basically babysit your deployment
Cue 419 — 00:27:23,728 --> 00:27:25,688
against the SLO target that you define.
Cue 420 — 00:27:25,728 --> 00:27:27,488
But we've struggled with adoption.
Cue 421 — 00:27:27,528 --> 00:27:31,608
And teams were saying like, my app is not critical enough, I don't need to apply
Cue 422 — 00:27:31,648 --> 00:27:37,488
this. So now we say that, hey, in order to reach the highest level of the readiness
Cue 423 — 00:27:37,528 --> 00:27:41,248
and to be able to be ready for autonomous changes, you have to actually adopt this
Cue 424 — 00:27:41,288 --> 00:27:42,388
platform feature.
Cue 425 — 00:27:43,728 --> 00:27:48,048
And having said that, I leave you with a quote that I frequently share with our
Cue 426 — 00:27:48,088 --> 00:27:52,208
engineers, especially those that are super excited by AI and forget that there is
Cue 427 — 00:27:52,248 --> 00:27:56,288
like a human judgment that needs to be there, which is not everything that can be
Cue 428 — 00:27:56,328 --> 00:28:01,928
built should be built. My name is Bartosz and you can find me here on socials.
Cue 429 — 00:28:01,968 --> 00:28:05,228
If you want to know more about, the tools that we have been building,
Cue 430 — 00:28:05,488 --> 00:28:07,928
here's the QR code that will lead you to our blog.
Cue 431 — 00:28:07,968 --> 00:28:08,988
Thanks a lot.
Cue 432 — 00:28:14,848 --> 00:28:18,368
We have time for 1 or 2 questions.
Cue 433 — 00:28:18,408 --> 00:28:20,608
If anyone's got a question, please raise your hand.
Cue 434 — 00:28:20,648 --> 00:28:22,668
I'll run a mic over to you.
Cue 435 — 00:28:24,808 --> 00:28:26,388
Had to be far away.
Cue 436 — 00:28:31,028 --> 00:28:36,868
Hey, thanks for the talk. When you were, like, giving us the outlook of the tools
Cue 437 — 00:28:36,908 --> 00:28:39,688
you have, it looks like a lot.
Cue 438 — 00:28:40,188 --> 00:28:42,688
And you mentioned that you started quite early.
Cue 439 — 00:28:43,228 --> 00:28:49,268
How do you judge whether this approach you've used two years ago or a year ago still
Cue 440 — 00:28:49,308 --> 00:28:55,048
stands where you need to rebuild from scratch and unlock some new capabilities.
Cue 441 — 00:28:57,948 --> 00:29:02,048
That have invested into one single tool like Claude or Cursor.
Cue 442 — 00:29:02,388 --> 00:29:05,268
They have been unable to integrate open weight models, right?
Cue 443 — 00:29:05,308 --> 00:29:08,668
They had to wait for the provider to provide it, and then providers are pricing them
Cue 444 — 00:29:08,708 --> 00:29:10,748
higher than they are available on the market.
Cue 445 — 00:29:10,788 --> 00:29:14,708
Where we are is we are able to integrate them on the proxy.
Cue 446 — 00:29:14,748 --> 00:29:16,708
And we already have developers who know the tools.
Cue 447 — 00:29:16,748 --> 00:29:20,388
We have the plugins on how you can work with OpenCode.
Cue 448 — 00:29:20,428 --> 00:29:24,148
And we also have a lot of developers who are advocates that OpenCode is actually
Cue 449 — 00:29:24,188 --> 00:29:26,468
giving them a lot of flexibility there, right?
Cue 450 — 00:29:26,508 --> 00:29:30,428
So this is basically our advantage because we are not bound to a single vendor.
Cue 451 — 00:29:30,468 --> 00:29:32,548
You mentioned one thing that it's like a lot.
Cue 452 — 00:29:32,588 --> 00:29:36,108
All those tools that I have shown are community built right.
Cue 453 — 00:29:36,148 --> 00:29:39,428
So the CLI, the plugins and so on and so on.
Cue 454 — 00:29:39,468 --> 00:29:43,068
This is like people who had an idea, people who were contributing.
Cue 455 — 00:29:43,108 --> 00:29:47,508
And my job is to create an ecosystem where they basically can safely contribute.
Cue 456 — 00:29:47,548 --> 00:29:53,388
If I see that a person is like very excited to, basically build a new version of the
Cue 457 — 00:29:53,428 --> 00:29:56,708
OpenCode plugin, they get contributor rights in the repository,
Cue 458 — 00:29:56,748 --> 00:30:00,028
and with code ownership they automatically get triggered on the PR.
Cue 459 — 00:30:00,068 --> 00:30:03,308
And this is what excites people because they get the visibility they get to help us
Cue 460 — 00:30:03,348 --> 00:30:04,368
on this journey.
.tessl-plugin
skills
talk-arcolano-autonomy-and-trust
talk-aroussi-decision-context
talk-bevin-agent-assisted-instrument-design
talk-boeckeler-context-lifecycles
talk-braun-context-graphs
talk-carmont-evolving-agent-harnesses
talk-chanezon-gordon-agent-experience-evaluation
talk-cox-secure-agentic-workflows
talk-cunniffe-reusable-agent-platform
talk-davis-agent-tool-policies
talk-debois-concurrent-delivery
talk-dzialets-team-automation
talk-fahmy-harness-api-design
talk-finster-continuous-delivery
talk-fowler-regenerative-software
talk-gavrilescu-factory-architecture
talk-graciolli-amin-verification-aware-programming
talk-gruenbaum-reverse-graph-engineering
talk-hassan-agentic-engineering
talk-hossain-enterprise-agent-adoption
talk-huntley-patch-ready-factories
talk-jain-verifying-intent
talk-katsioloudes-montei-agentic-lifecycle
talk-knox-harness-engineering
talk-kovac-deterministic-memory
talk-lee-perkins-code-readiness
talk-lorenc-machine-speed-security
talk-lubken-software-of-one
talk-meijer-software-specialization
talk-meppiel-versioned-agent-contracts
talk-moss-concurrent-agent-risk
talk-myers-understanding-agent-generated-code
talk-ocytko-scaling-agentic-engineering
talk-oza-automating-design-quality
talk-panzica-agentic-sdlc
talk-pari-patel-agent-awareness
talk-podjarny-trustworthy-software-factories
talk-poulter-factory-improvement
talk-pratik-patel-refined-rag-data
talk-re-ferre-development-platforms
talk-sadogursky-factory-of-factories
talk-shah-training-agent-skills
talk-sharma-mohapatra-evolving-intent
talk-singhal-operating-agents-at-scale
talk-stack-incident-dossiers
talk-tyler-team-coding-benchmarks
talk-wotherspoon-agent-execution-environments
talk-yegge-ai-powered-factories
talk-zuber-leadership-through-coding