---
格式版本: 2
标题: "Building Towards Self-Driving Codebases with Long-Running, Asynchronous Agents S81528 | GTC San Jose 2026"
原文链接: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/"
发布日期: "2026-08-14"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "Published Time: Fri, 14 Aug 2026 14:38:17 GMT"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 4
发布时间严格候选数量: 1
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-14T22:38:22+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-14T22:34:08+08:00"
入库时间: "2026-08-14T14:38:22.396Z"
来源平台: "固定入口"
搜索渠道: "fixed_url"
搜索词: "https://www.nvidia.com/gtc/"
匹配关键词:
- "performance"
- "latency"
- "throughput"
- "AI"
相关厂家:
- "NVIDIA"
- "OpenAI"
相关专家:
[]
内容类型: "网页"
抓取工具: "Jina Reader"
清洗工具: "Jina Reader Markdown + Defuddle/Readability 正文提取"
原始附件:
[]
AI优质: "否"
AI打分: 12
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "会议页面内容为AI编码代理（Cursor）演讲，与超节点/AI Rack/机柜级AI基础设施等硬件主题完全无关，属于无关资料。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-14T22:38:26+08:00"
AI主题相关性: 0
AI来源权威性: 10
AI新颖性: 2
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 0
AI摘要: "GTC San Jose 2026上，Cursor联合创始人Aman称AI编程正迈向异步代理，并介绍Cloud Agents：在云端VM长时运行、自动测试并产出视频工件；"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:30:35.659Z"
采集批次: "2026年8月14日18点30分41秒"
采集批次ID: "20260814-183041-680"
去重键: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528"
---
Title: Building Towards Self-Driving Codebases with Long-Running, Asynchronous Agents S81528 | GTC San Jose 2026
URL Source: https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/
Published Time: Fri, 14 Aug 2026 14:38:17 GMT
Markdown Content:
Visit your regional NVIDIA website for local content, pricing, and where to buy partners specific to your country.
[Continue](https://www.nvidia.com/)
[**GTC Berlin** October 20–22](https://www.nvidia.com/en-eu/gtc/) | [**GTC 2027** March 15–18](https://www.nvidia.com/gtc/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [](https://www.nvidia.com/en-us/account/)
* [Log In](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)[LogOut](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
PLATFORMS
other links
[](https://www.nvidia.com/gtc/)
Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
[Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/# "Menu")
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/# "Menu")
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81528/#)
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
Loading
Play
00:00
Play
Seek 10 seconds backwards
Seek 10 seconds forward
00:00 / 00:00
Mute
Press TAB to open volume control Use the arrows to control the volume
Turn on Picture in picture
Show Full screen
Transcript Powered by AI
X
X
00:11
Hey guys, I'm Aman.
00:12
I'm one of the founders of Cursor, and today I'm going to
00:15
be talking about what we think the future of AI coding will look like.
00:21
In particular, talking about async agents and this future
00:25
that we believe will look like self-driving codebases.
00:29
I'm going to start by walking through how coding has changed
00:32
over the last several months and years, and then go through
00:36
each of these future eras.
00:38
What will Async agents look like?
00:40
And what will self-driving codebases look like?
00:43
Finally, I'm going to end by talking through what the
00:46
role of the engineer looks like in a world where codebases
00:48
are fully self-driving.
00:52
So way back in 2021, 2022, the primary way that people were coding
00:58
was just using IDs with effectively no AI, no autocomplete even.
01:05
The only autocompletion was IntelliSense-level things.
01:09
But around that time, language models started getting much, much
01:13
better at coding, both generating code and understanding code.
01:18
And so the first AI coding application
01:21
that started to work was autocompletion, or tap-complete.
01:26
This is around the time that Cursor burst onto the scene.
01:29
And so roughly the way that this works is by looking at the last
01:34
however many minutes of work you've been doing inside of your editor.
01:38
And then, understanding that, models can predict what the next
01:43
few minutes of edits, places you're going to jump to, will look like.
01:46
And so it looks like this auto-completion, predicting the
01:49
next edits, and so on and so forth.
01:52
But models continue to get much better at code.
01:56
And they got better at being able to generate more complex changes,
02:01
potentially whole features.
02:04
And they got strong enough that you could ask with natural
02:07
language instructions for
02:10
Agents to be able to implement full changes.
02:12
And so that brings us to the next era, the era we're currently
02:15
in, where synchronous coding agents have quickly become
02:19
the dominant way of programming and coding with AI.
02:25
And you see this in the data.
02:28
2025 was really the year of coding agents completely taking off.
02:32
Where at the start of the year, the models become good enough and
02:36
the harnesses around them become good enough that you can actually
02:38
start to use coding agents for the first time for a lot of your work.
02:42
And then by the end of the year,
02:44
We're at a point where the vast, vast majority of code written
02:48
in Cursor is coming from agents rather than people using tab.
02:53
And this graph is actually astounding because it's not
02:55
just tracking users using agents versus tab, but it is
02:59
agent requests versus tab accepts.
03:02
And with tab, effectively every keystroke can potentially trigger
03:06
a tab request and get a request.
03:07
So you end up seeing
03:10
So many more people actually submitting whole prompts and
03:13
pressing enter than even writing keystrokes in the editor.
03:19
Now, synchronous agents work quite well, but there are
03:23
a lot of limitations.
03:24
You're mostly running these things on your local machines,
03:28
so they eat up a lot of resources.
03:30
And if we want to be able to write more and more code and kind
03:34
of increase the throughput, we're not going to be able to run tens
03:37
of agents on our local machine.
03:40
And so that brings us here.
03:44
We are in the second era of AI coding and slowly moving towards
03:49
this third era of async agents.
03:53
We believe async agents will need their own computers and
03:58
cloud environments to run.
04:00
If you want these agents to run for a very long time, they're gonna
04:03
need to be able to test their work.
04:06
And this means doing things like computer use or running
04:09
lots of tests that may be extremely resource intensive.
04:12
You can maybe scale this locally to a few agents, but scaling
04:16
to tens at a time is just not going to be possible.
04:20
And agents should have the full suite of tools
04:22
that a developer has.
04:25
And in these cloud environments, they should be able to do
04:27
whatever you could do locally.
04:32
And so here you can see the Cursor Cloud Agents product,
04:35
where we've given our agent access to this VM and a full desktop.
04:42
And it can use the desktop in the ways described before,
04:45
and one of the more interesting ways that it does use it is
04:48
via being able to use the computer and actually test out features.
04:53
I'll talk more about that later.
04:57
And internally, we've seen rapid adoption of Cloud Agents.
05:03
This is a chart that tracks until the end of February, and
05:07
it's quickly becoming the dominant way of coding within our company.
05:10
Around 30% of merged PRs are coming from Cloud Agents.
05:17
And there's one pretty... Actually, I'll talk about a few
05:21
examples of these PRs because it isn't just these trivial bug fixes.
05:25
One change made was there was a pretty big refactor
05:29
for video rendering that made things 25x faster and migrated
05:33
things from React to Rust.
05:36
And it took eight hours, and here the computer was extremely
05:39
useful for the agent to have because it could iterate on
05:43
making a change, actually running it and measuring the latency
05:46
of that, and then continuing to do it over and over again.
05:50
Another example was implementing this 10,000 line PR of adding
05:55
network policy controls for sandbox processes.
06:00
So what started to cause this inflection point?
06:03
Because you can see that around when it says here, artifacts
06:07
are released internally, you see this uptick in growth.
06:12
When we were using early versions of cloud agents, it started
06:16
to become really, really tricky to review all the changes
06:20
that agents were producing.
06:22
We're now operating in a regime where agents are producing
06:25
two, three, four X more code than they could have if just running
06:30
locally in a synchronous manner.
06:32
And reviewing all those code changes, and then iterating
06:35
with the agent when reviewing code manually just doesn't
06:40
become tractable at larger and larger levels of throughput.
06:45
And so in the same way that a manager might do a mix of
06:50
code review, but then also review the work of their engineer's
06:54
output, we think it's really important that when using
06:57
Cloud Agents, engineers can review the outputs of the model.
07:01
And we call these outputs artifacts.
07:05
One example of this is video artifacts.
07:08
So you can see here that after implementing a feature, we
07:11
have Cursor go in and actually take a video of the fully
07:16
functioning feature.
07:18
If there's a bug in how it implemented it, you can just
07:21
look at this video and then follow up with the agent.
07:24
Or if it implemented it in a way that didn't follow exactly
07:28
how you specified, similarly, you can re-prompt it without
07:32
having to look at the code to understand what's going on.
07:35
Then when you get to a working demo of what you think your
07:39
feature should look like, you can take a pass at the
07:41
code and review that and merge it in if things look good.
07:46
Similarly, for ML research, we found it pretty useful
07:51
for very small-scale experiments, having models go out and build
07:57
a research report.
07:58
So this is the output of what maybe a researcher might do.
08:01
And so you're not just reviewing the code changes
08:04
made by this agent.
08:07
But you're reviewing a report, which is far easier to grow
08:10
quickly and understand.
08:12
And again, it lets you iterate far more quickly and then
08:14
go to the code changes only when you have extreme confidence that
08:18
it's done the right set of things.
08:21
We're still figuring out what this looks like for places outside
08:25
of research and pure product.
08:28
A few additional ways that we think it could be helpful for backend
08:32
and infra might be iterating more so on things like architecture
08:36
diagrams, iterating more on plans.
08:39
It's possible that you want the models to over-test your code.
08:43
Thank you for watching.
08:44
Tests wouldn't be things that are actually checked into the codebase,
08:48
but they exist as artifacts that let you understand if the
08:52
agent did the right thing or not, and can be effectively thrown away.
08:57
But we haven't quite figured out what it looks like for
09:00
those other domains outside of product and research yet,
09:03
but we're working on it.
09:06
So how do we build Cloud Agents and make them extremely good
09:09
at long-running tasks?
09:12
Because this is the only way in which Cloud Agents and
09:16
Async Agents become useful, if you can actually trust
09:18
that they can run for a pretty long time and do a better job than
09:21
if they're running quickly locally.
09:25
So I want to introduce this concept of train time versus test time.
09:32
The way that these agents are trained, one of the main
09:36
ways in which these agents are trained is using RL, reinforcement
09:39
learning, and it will see a task in training, try to solve it many
09:44
different times, and then the times that it succeeded, those attempts
09:49
will get positively reinforced by the model, so the model will learn
09:52
to act more like that and learn to act less like the failure attempts.
09:57
And generally, if when you deploy your agent in the real
10:01
world, those problems that it sees, those tasks it sees,
10:05
are very similar to the tasks that it's been trained on, your agent's
10:09
going to perform a lot better.
10:10
So you want this train time, test time match, and if there's
10:14
a mismatch, you're going to see some degradation of performance.
10:21
So if you were to naively just have an agent work for a
10:26
very, very long time, it's going to have to run trajectories that span
10:31
not just hundreds of thousands of tokens, but millions, potentially
10:34
tens of millions of tokens.
10:36
And then this starts to get outside of the bounds of how these
10:40
models and agents are trained, where it is just intractable
10:44
to train in RL agents for things outside of hundreds of thousands,
10:49
maybe millions of tokens.
10:50
So as you go to high millions or tens of millions,
10:53
they'll just fall apart.
10:55
And you guys may have observed this from using codec agents
10:57
that run for extremely long times, but they'll lose track of
11:01
things, kind of not be able to go in enough detail in certain parts.
11:06
And this is where we find the concept of multi-agent
11:10
to be extremely useful.
11:12
Multi-agent is a broad term that effectively covers any system that
11:16
uses multiple agents in any way.
11:18
The simplest version of this is just the main agent and sub-agents.
11:23
And this is one of the key reasons why we believe sub-agents work
11:27
extremely well for longer running tasks, where each of these subtasks
11:32
is now a much simpler task that is well within the distribution
11:36
of how these models are trained.
11:39
And then the actual outer model that is calling all
11:43
these sub-agents is not running for as long a time, right?
11:46
Instead of running for tens of millions of tokens, it's running
11:49
for maybe a few hundred thousand.
11:51
And then calling sub-agents, kind of fanning out the work that way.
11:55
And so you get this nicer match where the models aren't kind
11:58
of overextended in how long they run versus how they're trained.
12:05
Another benefit of the multi-agent system is that you're starting
12:09
to see some kind of specialization or divergence in capabilities
12:14
and models for different things.
12:16
And what do I mean by this?
12:18
We found that OpenAI models tend to be the strongest when
12:21
it comes to higher level planning and orchestration.
12:25
But when it comes to things like computer use or multimodal
12:29
understanding, Gemini and anthropic models are stronger.
12:32
Or when it comes to creating better UIs, anthropic models
12:36
tend to be stronger.
12:38
And so the way that our cloud agents work is we
12:41
generally use OpenAI models as the planner, and then
12:44
we'll use other better multimodal models and computer-using models
12:49
to actually record the videos or use the features and prove
12:53
to you that things are working.
12:56
I think another interesting way that sub-agents can be
12:59
useful are, at a certain point, sub-agent tasks are not complicated
13:05
enough that you need the absolute largest, slowest models for it.
13:11
And you can use much faster models and get the exact same performance
13:14
for simpler sub-agent tasks.
13:17
And so we also end up doing this in the product and get the
13:19
same performance while delivering results much, much faster.
13:27
So what are the weaknesses of this system?
13:31
In the same way that the single agent running for a long time
13:35
would fall over at long task durations, this system also works.
13:40
But then as the task length also gets longer, eventually
13:46
it will hit its limit, where the planning steps and the outer agent
13:50
will start to go for millions, maybe tens of millions of tokens,
13:53
and then it will fall apart again.
13:56
And models
13:58
They still aren't trained super well to be excellent
14:02
orchestrators of agents.
14:04
And I think the way that RL works really benefits for
14:08
kind of the worker level tasks.
14:10
But there's a lot of work to do to make models much better at
14:12
this kind of orchestration layer.
14:15
It's something that we're very focused on when training
14:18
the next generation of our own models as well, especially
14:21
in being able to call out to multiple different models.
14:27
And I think another piece that is underrated is the
14:32
value of what we call Model UX.
14:35
And this is kind of the viewability of the model's outputs.
14:40
If we're starting to trust the artifacts produced by the
14:43
model more than the code itself and the code changes, then the model
14:49
needs to do a really good job of creating those artifacts for you.
14:52
In particular, if it recorded a video of a useless part of
14:57
the feature that it was building, or the part of the product
15:00
that it was building, it wouldn't be helpful at all, and you'd have
15:03
to go in and read the code anyways.
15:05
So models need to learn how to produce really useful and
15:08
understandable artifacts.
15:10
Already we see this at a kind of...
15:14
If different models produce the same markdown answer,
15:20
it's much easier to understand when it's nicely formatted.
15:23
And if it's this densely packed set of paragraphs, it's going
15:27
to take a lot longer to grok.
15:31
Understanding the actual thing that the model did, making
15:34
the model really good at doing that, will matter a lot, especially
15:39
as models get arbitrarily good at, given a well-defined
15:43
spec, producing a correct answer.
15:49
So now I want to talk about what we call self-driving codebases.
15:54
And we view this as somewhat of a kind of terminal state
16:00
of what async agents looks like.
16:05
And self-driving codebases, I think it involves a few details.
16:11
One is we want self-healing and fixing.
16:15
And this means that there are changes being made that
16:19
no human is ever reviewing.
16:22
Once more, as we go from local to async, you're increasing
16:26
the number of agents per person from a few to tens.
16:30
But at a certain point, if you want to increase your
16:32
productivity and have more and more done, it's going to be
16:35
too many agents running for people to review some agents' outputs.
16:40
And so you want some changes that just
16:42
automatically get into main and you can imagine this looks
16:45
like an issue tracker where issues are reported and some
16:49
issues are easy enough to fix and models and agents have high
16:53
enough confidence that they can go make the fix they can get into
16:57
main and no one's ever looked at it
17:01
Then the other piece is being able to significantly increase the
17:05
scope of what your agents can do.
17:08
In particular, building full projects or products or extremely
17:12
large features with little to no human intervention.
17:16
Think of like excellent products, like I think of Notion or
17:21
PowerPoint and whatnot.
17:23
Can you have models do something like build PowerPoint with
17:26
little to no human intervention?
17:27
It will take time to get there, but I think something like
17:30
this might be possible.
17:34
And in this world, we believe there's going
17:36
to be no human-written code.
17:40
So first, let's talk about self-healing and fixing.
17:46
One of our early attempts here at creating a product that can do
17:50
this is what we call automations.
17:54
And so you can trigger these things in certain events,
17:57
like the issue tracker case that I mentioned before, where
18:00
you can have agents on every issue proposing a potential fix.
18:05
But you can also have agents running at other
18:08
really valuable events.
18:10
Like every time you get paged in the middle of the night, we
18:15
have an agent going, investigating the issue, and then proposing
18:19
a potential solution.
18:20
So if you're waking up groggily at 2 or 3 AM, you might just need
18:25
a single click to fix the problem.
18:30
Eventually, we'd like to get to a spot where agents can
18:33
always be the primary on-call.
18:36
And there are issues that will pop up in the middle of
18:38
the night that an agent will fix, and no human will ever get paged.
18:43
Humans will just be escalated to as a secondary.
18:48
Already, we're using this in research.
18:52
For our training runs, we have agents running every
18:55
few steps, looking at the logs and the clusters, looking at weights
18:59
and biases to understand what's going on with all the metrics,
19:04
and then flagging potential issues.
19:06
This lets us catch potential failures in training earlier, so it
19:11
doesn't silently degrade the model.
19:13
And then it lets us catch issues that could lead to
19:16
runs crashing, preventing anyone from getting paged
19:19
or losing valuable training time.
19:24
We also do this for code review and for security, where we're
19:29
able to find a huge number of PRs with vulnerabilities that would
19:32
have otherwise gotten shipped.
19:34
So having these agents just always running at all these
19:36
events has already started to show serious signs of life.
19:44
Now, I want to talk about what that second part of
19:49
self-driving codebases looks like, meaning building full
19:52
projects, having really, really long-running agents with
19:56
little to no human intervention.
19:59
The browser was one of our attempts at this.
20:02
A browser is an extremely complicated piece of software, on
20:05
the level of or harder than an OS.
20:08
It requires a rendering engine that can take arbitrary HTML
20:12
and CSS and turn it into pixels, handling animations as well.
20:16
It requires being able to run arbitrary JavaScript in
20:19
the sandbox environment, and lots of other complicated things.
20:25
And so, it's a really daunting, challenging task, where the
20:28
ceiling can be quite high.
20:31
And we wanted to see if we could get agents to actually
20:35
build some kind of working browser.
20:38
This was a one-week run that took billions of tokens, on
20:43
the order of tens of thousands of dollars worth of compute,
20:46
and it produced something that works, but is still quite
20:49
far from real production browsers.
20:55
You can see an example of the browser here, where it's able
20:59
to render lots of arbitrary pages.
21:02
There are still several hiccups that don't make
21:07
it a fully functioning browser, but it's extremely impressive
21:11
that agents are able to do something like this today.
21:16
So how did we actually build a harness to make
21:20
something like this work?
21:22
It extended the harness that we were using for Async
21:25
agents and used the same concept of multi-agent.
21:30
So we don't want to let a given agent run for far too
21:33
long because it's just going to go off the rails.
21:37
And so what we do is we have a high-level planner that
21:40
can then call out to subplanners.
21:43
And it's this recursive-like thing where the subplanner
21:45
itself can call other subplanners.
21:48
And at the lowest level, or the leaf nodes, you have workers.
21:52
And we found this architecture.
21:54
We tried a bunch of different ones and this one ended up
21:57
working the best.
21:58
And it's a good bit simpler than kind of any of the other
22:02
ones we tried and really does leverage that same concept
22:06
of compressing the amount of tokens a given agent needs to run for.
22:12
The thing that we're pretty excited about is trying to
22:15
make this multi-model as well, where you can leverage those
22:19
same strengths of different models that we mentioned earlier,
22:23
where for the computer-using pieces, you'd use a different
22:25
set of models, as you would for planning and as you would for UI.
22:33
We're really excited about also training our models to
22:35
be quite good at these tasks.
22:39
I don't think any model has been trained with the idea
22:42
of being able to use it in this extremely long-running harness.
22:47
And we're pretty excited about training a model to act as
22:49
this planner or subplanner and be really good at calling
22:52
out to itself or other models.
22:59
So now we have this harness that works pretty well.
23:03
It's still early days and it'll take a lot of time to
23:06
get to something truly shippable.
23:09
But what does the right user interface look like in the
23:13
world where we have these self-driving code-based systems?
23:19
Right now, you have to do something that looks like
23:22
what prompt engineering was a year and a half or a few years ago
23:27
for getting these agents to work.
23:29
You have to write a really, really detailed spec, and
23:32
that goes into detail how it should think about breaking
23:34
down problems, really sets a clear rubric of what defines correctness,
23:41
and it just takes a lot of attempts and kind of nudging the harness
23:45
to get it to work really well.
23:46
So the UX isn't...
23:49
There yet we expect that as agents get better you'll be able to the
23:57
burden of how good the spec needs to be will decrease and you'll
24:02
be able to work with models as well on building out really good specs
24:06
for building these entire projects.
24:10
And then when you have agents building entire products,
24:16
what is the right level of review?
24:17
It would suck to do this many tens of thousands of dollar
24:22
run for a week and then come back and the output is garbage.
24:27
You wanna be able to intervene at subpoints in the middle.
24:30
I think it looks fairly similar to the same interfaces that
24:34
work for normal cloud agents, where you'll want the model
24:38
to be able to produce these artifacts in the middle, like
24:41
videos of various features.
24:44
The model is going to need to get much better at that
24:47
piece of model UX.
24:49
Because there's such a large surface area of what one could
24:52
test, what one could record, what one could try to show to the user.
24:56
And it needs to work on the areas of maximum ambiguity,
25:01
where the user didn't specify enough, or it's most unclear
25:05
what the right choice is, to have the user go in, to have the
25:08
engineer go in, and resolve that.
25:12
But this is still an open question, and we're working
25:15
on figuring that out.
25:19
So what does Cursor look like in this world, where you're
25:24
running these agents for extremely long time horizons, they're
25:27
tackling full products, full projects, and kind of automatically
25:33
healing your code base?
25:35
Well, in the same way that
25:38
The old generation of Clouds acts as what we
25:42
call a cost-of-good-sold Cloud.
25:46
In delivering your service, your product, to users, the old
25:51
Clouds deliver the infrastructure such that you can do this.
25:55
Cursor aims to be a new kind of Cloud that is building
26:00
an R&D Cloud that helps enterprises build more ambitious software.
26:07
And what does engineering become?
26:12
It's an interesting question, and we've
26:15
thought about this a good bit.
26:17
I'll preface it by saying, like, we don't have a fantastic answer, and
26:24
it's really hard to predict exactly what it'll look like, but...
26:29
The things that will be missing from agents and models for a while,
26:34
ironically, is kind of real agency of deciding what should get built.
26:38
There'll be details and product tastes that are
26:41
really hard to learn.
26:43
But ultimately, deciding what are the right things to build
26:46
that matter in the world are the most important pieces.
26:50
And this is why we're investing really heavily in growing
26:53
headcount ourselves.
26:55
We're scaling engineering 3x this year, and we don't
26:58
think that just because our engineers are getting more
27:01
productive with AI, we want to keep that same level of productivity but
27:05
cut headcount by a certain amount.
27:07
Instead, we want to tackle more ambitious things.
27:10
We want to increase our productivity and compound
27:13
really excellent engineers with excellent tools like Cursor.
27:19
Thanks, now I'll kick it over to the audience for a quick Q and A.
27:24
Excellent, any questions for Aman?
27:25
Yeah, oh.
27:31
Hi, I'm An.
27:32
My name is Michael Weald. I'm a student at Tulane University.
27:35
I have one semester left.
27:37
Just two weeks ago, I was actually taking a midterm in my class
27:42
for an artificial intelligence class and, well, over 90% of the
27:47
class, I want to say, was cheating.
27:50
Using, you know, artificial intelligence tools, large
27:52
language models to effectively answer questions about
27:56
Well, how does AI work and how can we build agents for
27:59
certain problems?
28:00
So I think it's illustrating that, like, as you said, the
28:05
role of writing programs is definitely changing.
28:09
How do you think Cursor, you know, perhaps aligns with
28:12
the vision of, say, what someone's learning in education?
28:16
And I don't know, what do you think, say, for instance,
28:19
someone that's studying computer science or studying to do
28:21
something like software engineer, how that's changing and how
28:24
Cursor is changing that?
28:28
Yeah, I think education adapts extremely slowly to these
28:36
rapid technological changes.
28:37
So I think lots of different organizations will be at different
28:41
places and how far along they are in trying to use these.
28:44
I think you kind of need to allow people to use AI.
28:49
For their applications. Just because that's not what things
28:52
will look like in the real world.
28:54
And today you definitely still need to really understand the
28:59
details of the code when operating in big company code bases.
29:02
But I think that will probably be less and less
29:05
the case in the future.
29:06
So I do think that organizations should try to, and educational
29:12
organizations should try to adapt the way that they teach
29:15
to account for this future.
29:18
Great question.
29:20
Way in the back.
29:31
Hi there, thanks for the talk, really appreciated it.
29:34
There's a lot of discourse around taste these days, you
29:37
were mentioning it as well.
29:40
It's interesting because people are also learning a lot from these
29:45
models and using them as well.
29:47
How do you guys think about both protecting taste as people
29:54
develop or cultivating taste as well in what you build?
30:00
I think it's really important to not lose to slop.
30:06
And you need to make sure that, like, people still design
30:13
and spend a lot of thought on what the shape of the UX
30:17
will look like and what the shape of the product will look like.
30:20
And making sure you don't just let velocity take over that.
30:24
Because I think taste matters in a few places.
30:27
There is the product side of making sure you're building
30:30
the right features, the right UX.
30:32
And then I think taste also matters in the architectural
30:35
and infrastructure side, and will matter for a while, of making
30:38
sure that you don't just merge slop code changes and really, really
30:43
bad architecture or code design.
30:47
So just making sure that you really value that as a company
30:51
and kind of state this clearly is important and just like
30:55
don't lose the battle to slop.
31:00
Question here. I'm on great talk.
31:02
Thank you. Regarding, you know, generating code is great,
31:06
but what about code reviews?
31:08
Like, do you have any thoughts on that?
31:11
Yeah, I mean, I think in the same way that models will
31:14
get much better at producing that new code, they're going to
31:17
get a lot better at reviewing code.
31:20
And reviewing code for humans is most of the time a lot
31:24
less fun than writing code.
31:27
And so now you have this interesting thing where
31:30
it's an activity that humans prefer not to do, and models
31:33
can be really, really great at.
31:37
I think we still need to get better, and it's really easy
31:41
to train models to be good at kind of writing code, given the way
31:44
that these training systems work.
31:46
I think we need to get better at figuring out how to do
31:49
that for code review.
31:50
Already they're decent, but I think there's just more work ahead
31:53
of us on the RL side in particular.
31:59
Hey, V from EAI.
32:02
You said you're scaling your engineering 3X, so as you're
32:05
assessing the folks you're looking to bring on board for that
32:08
agency skill and agent-building skills, like, how are you
32:12
assessing them, or what's your mechanism for figuring that out?
32:17
So the final stage of the interview process is a two-day,
32:22
one or two-day on-site project where they get to use Cursor.
32:26
So they get to use Agents as much as they want, and
32:28
the difficulty there is we've had
32:31
I've consistently increased the scope of what that project
32:36
demands as the products have gotten as agents have gotten
32:40
much, much better.
32:42
But we find that to be an excellent way of testing whether
32:45
people can use agents effectively and we test for the things that
32:49
matter, which are that agency piece and making the correct decisions
32:53
with respect to the architecture, the code or product taste or so on
32:56
and so forth for other disciplines.
33:00
Great question.
33:02
Hi, Aman. Thank you for the session.
33:03
I'm Suresh from United Health Group.
33:05
So definitely like the concept, the way you're explaining
33:08
the planners, the planner and then worker agents.
33:11
So have you thought about, if you look at the enterprise,
33:14
right, we still have large code bases sitting on legacy,
33:18
like AS400's main forms, right?
33:20
So it's very difficult to understand.
33:22
It's really what we really need agencies to understand
33:26
the concept and decompose, and right, even provide some kind of
33:30
a design even before start coding.
33:31
Yeah.
33:34
So for code understanding and I think there's two pieces there.
33:39
There's code understanding and then there's how do you actively
33:42
make your code base better.
33:45
One other part of a self-healing code base and self-driving
33:48
code base that I didn't mention is while you're sleeping
33:52
Uh, while you're not actively using these agents, they should be
33:55
actively improving the code base.
33:57
If there's tech debt or pieces of code that are really gnarly,
34:01
complicated, hard to understand, they should be going in and
34:03
just cleaning those up with the spare compute and off-peak hours.
34:09
And then, yeah, I think agents have been excellent
34:12
for understanding code.
34:14
A lot of the time, the way that people work for, do engineering
34:19
at Cursors, they spend a lot of time working with the agent
34:22
to understand some part of the code base before actually
34:24
writing a plan and then having the agent implement that.
34:28
So it's something we care a lot about.
34:31
And similarly for understanding harder parts of the code base,
34:35
we think, you know, spending more computers, scaling this thing up
34:37
in the multi-agent way could help.
34:40
Great question. I think we have time for two more questions.
34:43
So here you are. Hey, thanks for the presentation.
34:47
Just a quick question on the enterprise software realm, right?
34:53
As coding gets easier, agents are doing more coding.
34:56
Do you think there'll be a proliferation of
34:59
enterprise software?
35:01
Like there'll be more software companies, you know, more
35:04
agents of record, things like that?
35:07
Um, yeah, I think you're going to see it easier and easier
35:11
to create net new software, to maintain complex software,
35:16
and so I expect that...
35:18
You'll be more bottlenecked by the quality of your ideas and deciding
35:23
what to build than anything else.
35:25
And this allows a lot more companies to pop up, starters
35:28
to pop up, people to, and existing companies also to
35:32
build more products and enterprise, additional enterprise offerings.
35:36
So I see this kind of affecting all players in the space and
35:40
what they can do.
35:46
Hi, I'm Joe Felder, software engineer at NASDAQ, and I had
35:51
a question about the shift of the engineer's role and what you think
35:56
of this, that in the past, the best engineers can pay attention to
36:01
details, can hold a lot of things in their head, develop algorithms,
36:05
all that sort of thing, and that was a really important skill for
36:08
a good engineer, as well as, you know, top-down versus bottom-up,
36:12
looking at things and all that.
36:14
What do you think will be the next big skill for this
36:18
new way of doing things that an engineer would need to
36:21
have since they don't have to focus on the details that way anymore?
36:25
So I think, like, there are the skills that I had
36:29
mentioned earlier of taste and architecture and product.
36:34
But I think the same initial two things that
36:36
you mentioned still apply.
36:38
So holding a lot in your head is extremely valuable.
36:41
Now, just the bar is kind of increased or like the types of
36:44
things that you hold in your head.
36:45
maybe change so maybe you instead of an engineer holding
36:48
a lot of a particular part of the code base in their head an engineer
36:53
may need to hold the entire code base in their head but not need
36:56
enough detail of the lowest levels that they're holding in their
36:59
head so actually i think holding a lot in your head will still
37:02
be quite valuable to be effective and working with these agents
37:07
And then attention to detail.
37:09
I just think like the, what the details are will change,
37:14
but I do think that same attention to detail matters, right?
37:16
On the product side, models will still make, you know, mistakes in
37:20
various ways and misunderstanding what the user wants and will
37:24
do and have being able to pay attention to detail there.
37:27
Or maybe they just don't have sufficient context in making
37:29
the right architectural decisions.
37:32
And being able to pay attention there in what the model is
37:35
proposing to you and its outputs and the artifacts will still
37:38
be quite valuable.
37:41
Fantastic. I think those were some of the best questions
37:43
we've had in sessions all day.
37:45
So I want to take a moment and thank Aman so much for
37:47
a fantastic presentation.
37:50
Everyone put your hands together.
# Building Towards Self-Driving Codebases with Long-Running, Asynchronous Agents
Aman Sanger,CTO/Co-Founder,Cursor
Rate Now
Aman Sanger, co-founder and CTO at Cursor, will share how Cursor is building long-running coding agents that can autonomously execute more ambitious software tasks.
### Learn More About This Topic
Share
Favorite
Add to list
PDF
Events & Trainings:GTC San Jose
Date:March 2026
Topic:Agentic AI / Generative AI - Code / Software Generation
Region:
Company Information
* [About Us](https://www.nvidia.com/en-us/about-nvidia/)
* [Investors](https://investor.nvidia.com/home/default.aspx)
* [Venture Capital (NVentures)](https://www.nvidia.com/en-us/startups/nventures/)
* [NVIDIA Foundation](https://www.nvidia.com/en-us/foundation/)
* [Research](https://www.nvidia.com/en-us/research/)
* [Corporate Sustainability](https://www.nvidia.com/en-us/sustainability/)
* [Technologies](https://www.nvidia.com/en-us/technologies/)
* [Careers](https://www.nvidia.com/en-us/about-nvidia/careers/)
News and Events
* [Newsroom](https://nvidianews.nvidia.com/)
* [Company Blog](https://blogs.nvidia.com/)
* [Technical Blog](https://developer.nvidia.com/blog/)
* [Webinars](https://www.nvidia.com/en-us/about-nvidia/webinar-portal/)
* [Stay Informed](https://www.nvidia.com/en-us/preferences/email-signup/)
* [Events Calendar](https://www.nvidia.com/en-us/events/)
* [GTC AI Conference](https://www.nvidia.com/gtc/events/)
* [NVIDIA On-Demand](https://www.nvidia.com/en-us/on-demand/)
Popular Links
* [Developers](https://developer.nvidia.com/)
* [Partners](https://www.nvidia.com/en-us/about-nvidia/partners/)
* [Executive Insights](https://www.nvidia.com/en-us/executive-insights/)
* [Startups and VCs](https://www.nvidia.com/en-us/startups/)
* [Documentation](https://docs.nvidia.com/)
* [Technical Training](https://www.nvidia.com/en-us/learn/organizations/)
* [Professional Services for Data Science](https://www.nvidia.com/en-us/support/enterprise/advisory-services/)
Follow NVIDIA
[](https://www.facebook.com/NVIDIA "")[](https://www.instagram.com/nvidia/?hl=en)[](https://www.linkedin.com/company/nvidia/)[](https://twitter.com/nvidia "")[](https://www.youtube.com/user/nvidia)
[United States](https://www.nvidia.com/en-us/location-selector/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [Your Privacy Choices](https://www.nvidia.com/en-us/about-nvidia/privacy-center/)
* [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/)
* [Accessibility](https://www.nvidia.com/en-us/about-nvidia/accessibility/)
* [Corporate Policies](https://www.nvidia.com/en-us/about-nvidia/company-policies/)
* [Product Security](https://www.nvidia.com/en-us/product-security/)
* [Contact](https://www.nvidia.com/en-us/contact/)
Copyright © 2026 NVIDIA Corporation
Select Location
The Americas
* [Argentina](https://www.nvidia.com/es-la/ "Argentina")
* [Brasil (Brazil)](https://www.nvidia.com/pt-br/ "Brasil (Brazil)")
* [Canada](https://www.nvidia.com/en-us/ "Canada")
* [Chile](https://www.nvidia.com/es-la/ "Chile")
* [Colombia](https://www.nvidia.com/es-la/ "Colombia")
* [México (Mexico)](https://www.nvidia.com/es-la/ "México (Mexico)")
* [Peru](https://www.nvidia.com/es-la/ "Peru")
* [United States](https://www.nvidia.com/en-us/ "United States")
Europe
* [België (Belgium)](https://www.nvidia.com/nl-nl/ "België (Belgium)")
* [Belgique (Belgium)](https://www.nvidia.com/fr-be/ "Belgique (Belgium)")
* [Česká Republika (Czech Republic)](https://www.nvidia.com/cs-cz/ "Česká Republika (Czech Republic)")
* [Danmark (Denmark)](https://www.nvidia.com/da-dk/ "Danmark (Denmark)")
* [Deutschland (Germany)](https://www.nvidia.com/de-de/ "Deutschland (Germany)")
* [España (Spain)](https://www.nvidia.com/es-es/ "España (Spain)")
* [France](https://www.nvidia.com/fr-fr/ "France")
* [Italia (Italy)](https://www.nvidia.com/it-it/ "Italia (Italy)")
* [Nederland (Netherlands)](https://www.nvidia.com/nl-nl/ "Nederland (Netherlands)")
* [Norge (Norway)](https://www.nvidia.com/nb-no/ "Norge (Norway)")
* [Österreich (Austria)](https://www.nvidia.com/de-at/ "Österreich (Austria)")
* [Polska (Poland)](https://www.nvidia.com/pl-pl/ "Polska (Poland)")
* [România (Romania)](https://www.nvidia.com/ro-ro/ "România (Romania)")
* [Suomi (Finland)](https://www.nvidia.com/fi-fi/ "Suomi (Finland)")
* [Sverige (Sweden)](https://www.nvidia.com/sv-se/ "Sverige (Sweden)")
* [Türkiye (Turkey)](https://www.nvidia.com/tr-tr/ "Türkiye (Turkey)")
* [United Kingdom](https://www.nvidia.com/en-gb/ "United Kingdom")
* [Rest of Europe](https://www.nvidia.com/en-eu/ "Rest of Europe")
Asia
* [Australia](https://www.nvidia.com/en-au/ "Australia")
* [中国大陆 (Mainland China)](https://www.nvidia.com/zh-cn/ "中国大陆 (Mainland China)")
* [India](https://www.nvidia.com/en-in/ "India")
* [日本 (Japan)](https://www.nvidia.com/ja-jp/ "日本 (Japan)")
* [대한민국 (South Korea)](https://www.nvidia.com/ko-kr/ "대한민국 (South Korea)")
* [Singapore](https://www.nvidia.com/en-sg/ "Singapore")
* [台灣 (Taiwan)](https://www.nvidia.com/zh-tw/ "台灣 (Taiwan)")
Middle East
* [Middle East](https://www.nvidia.com/en-me/ "Middle East")
NVIDIA uses cookies to improve your experience on our web site. We and our third-party partners also use cookies and other tools to collect and record information you provide as well as information about your interactions with our websites for performance improvement, analytics, and to assist in marketing efforts. By clicking "Accept All", you consent to our use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/). You can manage your cookie settings by clicking on "Manage Settings." By continuing to use this site or by clicking one of the buttons below, you agree to our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control (GPC) signal and recorded your rejection of all optional cookies on this site for this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. To opt out of non-cookie personal information "sales" / "sharing" for targeted advertising purposes, please visit the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies. You can manage these settings in the NVIDIA [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies which overrides at least one of your previous settings. You can manage them in the [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
Manage Settings
Reject Optional Accept All
Cookie Settings
We and our third-party partners (including social media, advertising, and analytics partners) use cookies and other tracking technologies to collect, store, monitor, and process certain information about you when you visit our website. The information collected might relate to you, your preferences, or your device. We use that information to make the site work, analyze performance and traffic on our website, provide a more personalized web experience, and assist in our marketing efforts.
Under certain privacy laws, you have the right to direct us not to "sell" or "share" your personal information for targeted advertising. To opt-out of the "sale" and "sharing" of personal information through cookies, you must opt-out of optional cookies using the toggles below. To opt out of the "sale" and "sharing" of data collected by other means (e.g., online forms) you must also update your data sharing preferences through the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/).
Click on the different category headings below to find out more and change the settings according to your preference. You cannot opt out of Required Cookies as they are deployed to ensure the proper functioning of our website (such as prompting the cookie banner and remembering your settings, etc.). By clicking "Save and Accept" or "Decline All" at the bottom, you consent to the use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) in accordance with your settings and accept our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). For more information about our privacy practices, please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/).
Required Cookies
Always Active
These cookies enable core functionality such as security, network management, and accessibility. These cookies are required for the site to function and cannot be turned off.
Cookies Details
Performance Cookies
- [x] Performance Cookies
These cookies are used to provide quantitative measures of our website visitors, such as the number of times you visit, time on page, your mouse movements, scrolling, clicks and keystroke activity on the websites; other browsing, search, or product research behavior; and what brought you to our site. These cookies may store a unique ID so that our system will remember you when you return. Information collected with these cookies is used to measure and find ways to improve website performance.
Cookies Details
Personalization Cookies
- [x] Personalization Cookies
These cookies collect data about how you have interacted with our website to help us improve your web experience, such as which pages you have visited. These cookies may store a unique ID so that our system will remember you when you return. They may be set by us or by third party providers whose services we have added to our pages. These cookies enable us to provide enhanced website functionality and personalization as well as make the marketing messages we send to you more relevant to your interests. If you do not allow these cookies, then some or all of these services may not function properly.
Cookies Details
Advertising Cookies
- [x] Advertising Cookies
These cookies record your visit to our websites, the pages you have visited and the links you have followed to influence the advertisements that you see on other websites. These cookies and the information they collect may be managed by other companies, including our advertising partners, and may be used to build a profile of your interests and show you relevant advertising on other sites. We and our advertising partners will use this information to make our websites and the advertising displayed on it, more relevant to your interests.
Cookies Details
Cookie List
Clear
* - [x] checkbox label label
Apply Cancel
Consent Leg.Interest
- [x] checkbox label label
- [x] checkbox label label
- [x] checkbox label label
Decline All Save and Accept
[](https://www.onetrust.com/solutions/consent-and-preferences/)
Copy debug info