[00:13] Welcome everyone.
[00:15] Wow, this is a big room.
[00:18] We're so excited to be here today and
[00:20] share what's new in Agent Platform and
[00:22] do a deeper dive on what we just
[00:24] announced in the keynote and show a
[00:27] little bit about some how some of our
[00:28] customers are adopting the Agent
[00:30] Platform already.
[00:32] My name's Aman Khan. I'm a group product
[00:34] manager on the Agent Platform team
[00:36] focusing on agent governance. And I'm
[00:38] joined by Lavi on our DevRel team who's
[00:40] going to be doing a demo of some of what
[00:42] we talked about earlier today as well as
[00:44] some amazing partners from L'Oréal,
[00:47] Gautier and Thomas who are going to be
[00:48] showing how they're already adopting the
[00:50] Agent Platform for production workflows
[00:53] today.
[00:55] So, I want us to zoom out really quick
[00:57] before we jump into all of this and
[00:59] think about how exactly 1 year ago
[01:02] you learned about ADK. We showed you all
[01:05] Agent Engine. We wowed you with a bunch
[01:07] of demos like the autonomous plant store
[01:10] and we're really excited about what was
[01:12] really possible with the Agentic AI
[01:14] cloud that we launched last year.
[01:16] But since then, we've actually made a
[01:18] number of improvements like in November
[01:21] when we launched the basics of additions
[01:23] of governance
[01:25] and the basics of deployment to allow
[01:27] you to go from prototype to production.
[01:31] And it feels like things are changing
[01:33] again, right? This should feel familiar.
[01:35] Your competitors are moving faster. The
[01:39] industry is moving at an exponential
[01:41] rate.
[01:42] And technology isn't moving linear
[01:44] anymore. It's advancing exponentially.
[01:46] But everyone's organization is still in
[01:48] this sort of logarithmic trend. The
[01:50] curve sort of flattens out, right? Think
[01:53] organizational challenges, compliance,
[01:56] security, getting your whole org bought
[01:58] in on a single stack.
[02:00] That's why you have to think about step
[02:02] functions in how you operate.
[02:06] And that's what we're going to be
[02:07] talking about. How the Agent Platform
[02:09] enables these step function improvements
[02:11] to how you can adopt Agentic AI
[02:13] applications.
[02:15] So, a year ago everything was sort of
[02:17] centered around chatbots, right? Like
[02:20] rag chat applications.
[02:22] Every demo was a chatbot.
[02:24] But now it's pretty clear that agents
[02:27] are the engine of our business.
[02:29] Think of it this way. I heard this
[02:31] analogy earlier this week and it's
[02:33] really stuck with me.
[02:35] Everyone can get in a Toyota Camry and
[02:37] drive it. It's a reliable car. It's
[02:39] stable, runs well. That's traditional
[02:42] software development. But with agents,
[02:45] we're building race cars. We're building
[02:46] Formula 1 race cars. These are highly
[02:49] sophisticated instruments that can do
[02:51] real work in your business.
[02:53] But you can't just run a race car in
[02:55] yolo mode. You wouldn't put an electric
[02:57] car in self-driving mode and just floor
[02:59] it, right? Actually, quick question for
[03:01] folks in the room. How many people have
[03:03] actually tried Open Claw and built
[03:05] something with Open Claw?
[03:07] How many folks have heard of Open Claw
[03:09] at least? Okay, there we go.
[03:11] Think of it this way.
[03:13] What would you say if someone came up to
[03:15] you and said, "Hey, maybe we should put
[03:16] Open Claw in our business and let it
[03:19] start running and see what happens?"
[03:22] Yeah. That's kind of a crazy
[03:24] proposition.
[03:26] But what if it wasn't that crazy? What
[03:28] if it actually worked?
[03:31] Well, you need guardrails to make sure
[03:32] that when you're building agents, you're
[03:34] doing it in a way that's integrated with
[03:36] an end-to-end stack with context in mind
[03:39] from the start. And that's what we're
[03:41] here to do today. We're going to show
[03:43] you how to have an end-to-end ability
[03:45] with building agents using the Agent
[03:47] Platform.
[03:49] To enable this shift, we're combining
[03:51] Gemini Enterprise with the Agent
[03:53] Platform in one story, one narrative,
[03:56] and one seamless experience.
[03:58] Gemini Enterprise is the doorstep for
[04:00] everyone in your organization to get
[04:03] started with AI
[04:05] and Agent Platform is where your
[04:07] developers build.
[04:08] What this means is that your business's
[04:11] processes stop being disconnected.
[04:13] Agents and data can actually flow
[04:16] through your enterprise without
[04:17] friction.
[04:18] There's no handoff tax between what your
[04:20] developers are building on the back end
[04:22] and what your employees and customers
[04:25] actually use.
[04:26] At the top,
[04:28] the surfaces that your users interact
[04:30] with like the Gemini app, custom apps,
[04:32] things that partners may build. And then
[04:34] in the middle, the Agent Platform which
[04:36] is where your developers actually live.
[04:38] The runtime, registry, gateway,
[04:41] identity, observability, and evals.
[04:43] All of that sits on top of Google
[04:45] Cloud's AI stack, the hypercomputer,
[04:48] Agentic Data Cloud, and Agent Defense.
[04:51] This foundation is what gives you a real
[04:54] edge in performance and security because
[04:56] you don't need to stitch together five
[04:58] different vendors to get this. Now,
[05:00] let's zoom into the Agent Platform
[05:02] layer.
[05:03] Agent Platform is a vertically
[05:05] integrated stack that lets you move from
[05:07] basic idea to a production-ready agent
[05:10] without having to rebuild the underlying
[05:12] plumbing every single time you start a
[05:14] new project.
[05:16] With Agent Platform, you get
[05:17] best-in-class support with open-source
[05:20] native tooling to build agents the way
[05:22] you want to build them.
[05:24] Our goal centers on three areas. First,
[05:27] deeper investment in a richer richer
[05:29] ecosystem and interoperability.
[05:32] Second, ease of use to get started so
[05:35] that your entire team can collaborate in
[05:37] one place. And third, customizability
[05:40] because we know that off-the-shelf
[05:41] agents don't just work. You have to hook
[05:44] them up to your data, your tools, and
[05:46] run within the bounds of your business.
[05:49] So, that's the core foundation for your
[05:51] developers, your AI engineers, your AI
[05:53] product managers to get started with
[05:55] building agents. Agent Platform lets you
[05:57] build, scale, govern, and optimize
[06:00] agents that can scale infinitely in the
[06:02] cloud.
[06:03] Now, we're going to get into the how.
[06:04] We'll spend time on each part of the
[06:06] stack, what's new and why it matters,
[06:09] and then we'll talk you through this
[06:10] from the perspective of a developer
[06:11] that's planning agents in your business
[06:14] this year. So, I want you to kind of put
[06:15] yourself in that mindset, right? Like,
[06:17] what are the agents that you're thinking
[06:19] about in this next half that can really
[06:21] move the needle for your business? And
[06:24] let's see what it would take to actually
[06:26] ship that with the Agent Platform. If
[06:28] there's any parts that really stand out,
[06:30] come and find one of us for a deeper
[06:31] dive after this presentation.
[06:34] Let's start from the beginning,
[06:35] building.
[06:37] ADK has been our foundation to help you
[06:39] get started with building and this year
[06:41] we've made some major advancements to
[06:42] make it even more powerful and
[06:44] extensible as an agent harness
[06:46] framework. Google uses ADK widely
[06:49] throughout the business from bets in
[06:51] different parts of Alphabet to core
[06:53] products and services you've probably
[06:55] interacted with just today.
[06:57] It's designed to help you scale from the
[06:59] smallest agent to the most complex
[07:01] multi-agent system and to enable
[07:04] interoperability.
[07:06] A few exciting launches where we're
[07:08] excited to announce. Graph-based
[07:09] orchestration means you can model real
[07:12] business processes, real work branches
[07:14] and loops and conditional paths.
[07:17] You have agent collaboration which means
[07:19] you can decompose complex problems
[07:21] across specialized agents instead of
[07:23] trying to cram everything into one
[07:25] single monolithic agent. And then you
[07:27] have batch and event-driven workflows so
[07:29] that agents can actually react to things
[07:31] in real time that are happening in your
[07:33] system versus just reacting when a human
[07:35] tells them to.
[07:36] We also are excited to share that we
[07:38] have native skill definitions in ADK
[07:41] which means you can define a skill once
[07:43] and reuse it everywhere else ADK is
[07:46] being used in your business.
[07:49] Getting started is also about to get a
[07:51] whole lot easier.
[07:52] We're excited to launch a new agent CLI
[07:55] and skills for building, evaluating, and
[07:58] deploying AI agents on Google Cloud.
[08:01] Agent CLI is built with ADK and our
[08:04] agent runtime in mind so that that
[08:07] knowledge of our entire platform is
[08:09] actually baked into these skills and
[08:11] CLIs that you're getting to use. So, if
[08:12] you're just getting started with
[08:14] building an agent, the CLI can handle
[08:16] everything around it like the
[08:17] scaffolding, evaluation, deployment, and
[08:19] observability. And it works in two ways.
[08:22] You can either just pass the CLI and
[08:24] skills to the coding agents that you're
[08:26] already using like Cloud Code, Gemini
[08:29] CLI, Code Acts, or others and the coding
[08:31] agent can make the right decisions. Or
[08:34] if you're not using a coding agent, you
[08:35] can just run the CLI commands directly
[08:37] from your terminal and every command
[08:39] works standalone.
[08:40] Personally, I found it incredibly
[08:42] helpful to help get my agents stood up
[08:44] in our environment and we're just
[08:46] getting started with out-of-the-box
[08:47] primitives for building on top of the
[08:49] platform.
[08:50] Next, let's talk about scale.
[08:53] Building agents is step one, but running
[08:55] agents in confidence in production is
[08:58] the critical next step. This starts with
[09:01] deploying your agent to a runtime that
[09:03] enables you to serve thousands or even
[09:05] millions of users right out of the box.
[09:08] With Agent Platform, we're making the
[09:10] runtime significantly more powerful with
[09:13] sub-second cold starts. And we've added
[09:15] the capability to bring your own
[09:17] container for ultimate flexibility for
[09:19] how you want to run those agents.
[09:21] We also know that agents don't just run
[09:22] out of the box on a task for a long
[09:24] duration without some serious
[09:26] configuration which is why we're excited
[09:28] to share that the Agent Engine now also
[09:30] supports individual agents to be able to
[09:33] run on a single plan without any human
[09:36] intervention for up to 7 days. And
[09:38] that's alongside our already supported
[09:41] advanced multi-agent orchestration.
[09:44] Plus, we're adding bidirectional
[09:46] streaming to handle more complex
[09:48] real-time interactions like that Gemini
[09:50] flash demo you saw earlier for
[09:52] low-latency streaming interactions.
[09:56] Agents shouldn't start from scratch with
[09:58] every conversation.
[10:00] Without memory, your agent has no idea
[10:02] what happened yesterday, what the user
[10:04] prefers, what failed last time. And now,
[10:07] with sessions and memory bank, agents
[10:09] get persistent memory so that they get
[10:11] smarter over time by retaining contacts,
[10:14] recalling preferences, and learning over
[10:17] time so that you can build an
[10:18] intelligence layer that compounds.
[10:21] The session service is also
[10:22] framework-agnostic, so you can define
[10:24] custom session fields to capture signals
[10:26] like user feedback.
[10:29] Agents are most powerful when they can
[10:30] act in the real world, so we're excited
[10:33] to share a number of sandbox features in
[10:35] GA.
[10:37] Code execution sandboxes lets agents run
[10:39] in the real world world on real-world
[10:42] tasks, while bringing your own container
[10:44] support allows you to define and use
[10:46] custom environments when you have strict
[10:48] requirements. Agents can now interact
[10:50] with GUIs so that they can use the the
[10:52] computer use sandbox as well.
[10:55] Contexts makes agents smarter, and
[10:57] sandboxes make agents more useful.
[11:00] Together, you go from a chatbot that
[11:02] answers questions to an autonomous
[11:04] system that remembers what it's supposed
[11:06] to be doing, how to do it, and can
[11:08] actually go and execute tasks.
[11:12] Next, let's talk about governance.
[11:14] We're thinking about governance as an
[11:16] end-to-end part of our stack. Your
[11:18] governance story shouldn't be pieces
[11:20] that you're going to pull together from
[11:21] different vendors or legacy services you
[11:23] already have. You really need to think
[11:25] about governance as a critical part of
[11:28] the stack that works through the runtime
[11:30] all the way through to optimization,
[11:32] evals, observability.
[11:34] Agent platform provides that foundation
[11:36] for trust for every agent that you
[11:38] deploy. And we're excited to share that
[11:41] the agent identity is now in GA. You can
[11:44] manage permissions just like you would
[11:46] any other user in your organization with
[11:48] granular controls to ensure every action
[11:51] is authorized and auditable.
[11:53] Whether you're chief compliance officer
[11:55] or a risk officer or just a developer,
[11:57] you've probably asked yourself, "What
[11:59] are all of these different agents that I
[12:01] have across my business in the first
[12:02] place?"
[12:04] Agent registry helps you answer that
[12:06] question by listing all the agents that
[12:08] have been registered automatically via
[12:10] our agent runtime or the agent registry
[12:13] system.
[12:14] When agents are registered, you're able
[12:16] to see what tools those agents have
[12:18] access to, what services they can call,
[12:21] and it's just really this is a great
[12:22] starting point to really see how agents
[12:24] are being used across your enterprise.
[12:27] On top of that, we've made some serious
[12:29] enhancements to how you can secure your
[12:31] agents in your environment using model
[12:33] armor, which can block prompt injection,
[12:36] prevent data leaks in real time. You
[12:39] need to be able to monitor your agents'
[12:40] interactions and actions, and that's why
[12:42] we've launched proactive threat
[12:44] management directly integrated within
[12:46] the platform. And we're also launching
[12:48] agent gateway, which is a control plane
[12:51] for enterprise agents.
[12:52] The agent gateway simplifies how you
[12:54] secure and govern all of your agent
[12:56] traffic so you can set granular controls
[12:59] like access to specific tools. And if
[13:01] you need something custom, it's fully
[13:03] extensible. So, we're really thinking
[13:05] about governance as an end-to-end story.
[13:07] It's not bolted on as an afterthought.
[13:09] We think governance should span from
[13:11] prototype to production, and it includes
[13:13] the perimeters and boundaries to define
[13:15] what your agent can and cannot do in
[13:17] your environment. Today, we're launching
[13:19] those foundational components with
[13:21] batteries included, but our vision is
[13:23] that governance is architecturally
[13:24] integrated throughout the platform. And
[13:26] our goal is to build features so that
[13:28] the users of our platform don't need to
[13:30] be experts in governance to ship agents
[13:32] that your organizations can trust. It's
[13:34] really we think of it as like an agent
[13:36] alignment problem.
[13:38] So, we're excited to launch this
[13:39] foundation, and we're going to use it to
[13:40] build out a full governance story going
[13:42] forward.
[13:43] All right. We're rounding out here.
[13:44] We're going to get into this last
[13:46] pillar, optimization. Now, when it comes
[13:49] to optimization, here's what makes our
[13:51] approach different. You get the benefit
[13:53] of the research of DeepMind and how they
[13:56] build evals with a specialized set of
[13:58] tools for online and offline evals and
[14:01] continuous monitoring without having to
[14:03] have an army of researchers inside your
[14:05] own business. We provide you with that
[14:08] army of researchers throughout the agent
[14:10] platform because Google has invested
[14:12] deeply in how to evaluate our own
[14:14] agents, and you're getting the benefits
[14:16] of that investment because it's the same
[14:17] product.
[14:19] We're shipping auto raters, which are
[14:21] out-of-the-box scorers for offline and
[14:23] online evals, as well as an evaluation
[14:26] expert skill to make evaluation with the
[14:28] coding agent even simpler by passing
[14:30] them that skill and saying, "Hey, set up
[14:32] evals for me in the system."
[14:34] We're also shipping ADK optimization,
[14:36] which is based on Jeppa prompt
[14:38] optimization.
[14:39] With agent platform optimization, you're
[14:42] getting a single pane of glass because
[14:44] your whole team is already in the agent
[14:46] platform.
[14:47] The things slowing down your agents
[14:49] today aren't the model, it's tool calls,
[14:51] it's contacts. And when you build,
[14:53] govern, deploy, and optimize all in one
[14:56] system, you get a continuous feedback
[14:58] loop
[14:59] to really improve the agent over time.
[15:02] That means you get production quality
[15:04] agents that get better over time without
[15:06] needing to stitch together a bunch of
[15:08] different vendors and tools to get this
[15:10] out of the box.
[15:12] All right. That was a lot of ground, and
[15:14] with that, I'd like to welcome up Lavi
[15:16] to the stage, and he's going to take us
[15:17] through a demo of the agent platform.
[15:23] >> [applause]
[15:29] >> All right, let's switch to the demo.
[15:33] Perfect. So, Aman did talk about a lot
[15:35] of things. What we're going to do is
[15:37] we're going to see live what is
[15:38] available to you right now in your cloud
[15:41] console. I'm going to go to one of my
[15:43] projects,
[15:45] and this is the agent studio. So, this
[15:46] is the new studio that you see. So,
[15:48] Vertex AI is now sort of called as agent
[15:50] studio, where you have agents, you have
[15:52] app builder, and rate media. As you can
[15:54] see, you can actually add agents here
[15:56] directly, whatever agents that you have
[15:58] built. So, I'll start with this.
[16:00] The next thing I want to show is that
[16:02] you can actually create your new agents
[16:03] right away. So, the moment I do this,
[16:07] you can see that there's a UI now, and
[16:09] you don't have to write code if you
[16:10] don't want to, but if you wish to, you
[16:12] can actually just do the UI designer
[16:13] part of that. What we also have is bunch
[16:17] of different agents which are available
[16:19] as agent garden. So, if you have already
[16:21] used model garden, you know that what it
[16:23] allows you to do is access different
[16:25] kinds of models beyond Gemini. Very
[16:27] similarly, agent garden gives you access
[16:29] to different kinds of agents which are
[16:30] either built by Google or somebody
[16:33] outside of that. So, you can actually
[16:35] choose bunch of these different agents
[16:37] right away, and then sort of build over
[16:38] that or just use it as is. As you can
[16:41] see, we have so many different options
[16:43] available here, which you can use and
[16:45] get started. We also have something
[16:46] which is ADK 2.0, like Aman was
[16:48] mentioning that we have introduced many
[16:50] of the new features, especially the
[16:51] graph-based workflows that you can
[16:54] actually now build in ADK. So, you can
[16:56] actually use that and see how you can
[16:58] build over this.
[17:02] Oops.
[17:03] Sorry. So, the next is the ADK. So, this
[17:06] is where you actually again get a lot of
[17:08] access to ADK docs right in your
[17:11] platform. And this is where you should
[17:13] be able to start with any of the
[17:15] documentations which are required. The
[17:17] other thing which is very important, and
[17:18] I'll I'll come back to this
[17:20] at the other stage, which is at the the
[17:22] agent registry part, but you now have a
[17:25] centralized access where you can
[17:26] actually see all the MCP servers that
[17:29] you have, whether it's your server
[17:30] outside, and you can sort of have a
[17:32] central location where you can see this.
[17:34] You also get vector search and rag
[17:36] engine. So, this is again something that
[17:37] we had before, but now it sort of all
[17:39] comes under the build stage. So, you
[17:41] have all of these accessible right to
[17:43] you, and you can sort of define your
[17:45] vector search, you can create new
[17:46] indexes if you wish to, all the options
[17:48] that are available to you. You can also
[17:50] create rag engine if you wish to do
[17:52] that, and you can sort of create corpus
[17:54] around it.
[17:55] And then the last one is the search. So,
[17:57] search is very important because again
[17:59] lot of
[18:00] search that you want to do, a lot of
[18:01] agents that you want to build, you need
[18:03] inbuilt capability for them to search
[18:05] through your documents. And that's what
[18:07] search sort of gives you access to,
[18:08] which is you can create a custom search
[18:10] or you can sort of do a media search
[18:13] as per that.
[18:14] The next thing which is very important,
[18:16] this is one of my favorite, which is the
[18:18] agent engine part of it. So, once you're
[18:20] done building, whether by designer or by
[18:23] code, one of the very important thing
[18:25] that you want to do is you want to use
[18:27] the agent engine to scale your agents,
[18:28] right? And we have new runtimes which
[18:30] are there. The way this looks like is
[18:32] that you have all the agents that you
[18:34] see here. So, these are bunch of agents
[18:36] that I was experimenting, and you can
[18:37] see those agents here.
[18:40] Once I click on this,
[18:42] the best part you can actually see is
[18:45] Let me sort of select this.
[18:48] Oops, sorry.
[18:59] I'm going to select one of the old
[19:01] agents that I was playing with, and this
[19:03] is where you can actually see bunch of
[19:06] metrics. So, for example, you get the
[19:07] overall view as to how many sessions
[19:09] were there. You see the evaluations, you
[19:11] see how many model calls were made. You
[19:14] also see the tools that are there. For
[19:16] example, I did close to 100 model calls
[19:19] on this. You also get to see the tools.
[19:22] So, this is again all the tools calls
[19:23] that happen in your agent,
[19:25] and the overall usage as well. So, this
[19:27] is again very important because you need
[19:29] that visibility into what your agent has
[19:32] done, and this sort of
[19:34] deployment sort of section gives you the
[19:35] complete view of that.
[19:37] The other very important things are the
[19:39] traces. Traces are very important
[19:41] because if something goes wrong in your
[19:43] agent, you need to have a view of that.
[19:46] The way that happens is if you're using
[19:47] ADK, you have session IDs, and these
[19:49] session IDs are very important. So, the
[19:51] moment you click on this, you actually
[19:53] get to see all the sessions and all that
[19:55] has happened. And you can then
[19:57] individually go and see the trace view
[19:59] or the span view sort of just to get to
[20:02] the part where things have not worked
[20:05] out.
[20:05] The other important thing is you have
[20:07] access to sessions. So, if your users
[20:09] are using and you have like thousands of
[20:11] sessions, you can actually see all
[20:12] individual sessions because now they
[20:14] have IDs which are exposed to you. The
[20:16] best part is you also have a playground.
[20:18] So, it's not that you just do a local
[20:20] testing. You actually have a way where
[20:22] you can
[20:22] test your agents right away in the
[20:25] playground. So, just to show you a
[20:26] couple of examples,
[20:28] if you open, you can actually see all
[20:30] the stuff that somebody has been playing
[20:32] or you have played with this in that
[20:34] sense. It also has memories now. So, if
[20:37] you have written your own memories or
[20:39] you have written skills or logic which
[20:40] can extract these memories, you will see
[20:42] that here as well.
[20:44] Going to the next which is the memory
[20:46] bank. This is again very important
[20:48] because you do want to have memories in
[20:51] your agents in terms of things which the
[20:53] user wants you to remember.
[20:56] And now you have a centralized place
[20:57] where you can store all of these things
[20:59] in a simple simple way.
[21:01] Again, session I've shown you which is
[21:03] very simple all the access to the
[21:06] sessions. The next thing which is very
[21:08] important is the governance part which
[21:10] is where you have access to all the
[21:12] agent registry. So, this is where I
[21:13] wanted to show you that now you have a
[21:15] centralized view where you have access
[21:18] to all the agents which are available in
[21:20] the agent platform. You also have all
[21:22] the MCP servers. So, at least you now
[21:24] get to have a centralized place where
[21:26] you can govern them. The most important
[21:28] part is that you get to see how many
[21:30] tools are registered inside it and
[21:32] there's a specific ID which is
[21:33] associated which means you can sort of
[21:35] create a governance chain around that.
[21:38] Similarly, you have endpoints and you
[21:40] can sort of add these MCP servers and
[21:42] all of those things. Same with policies.
[21:44] Again, because now you have all of this
[21:46] as a central ID level which means that
[21:48] you can actually add new policies and
[21:50] you can define the access by the users
[21:54] or by the organization at the higher
[21:56] level. Similarly, you have gateways.
[21:58] Gateways are very important because
[21:59] again, you would want to restrict
[22:01] certain things for certain users. You do
[22:03] not want them to access certain things
[22:04] which are not required. So, you can
[22:07] define different gateways. You can
[22:08] define different kinds of access path.
[22:11] You can enforce policies. You can audit
[22:13] policies and all of those things. One of
[22:15] the cool things that you can do now is
[22:17] you can enable the model armor by
[22:19] default. You can sort of select the
[22:20] agent and then you can enable those
[22:22] model armors things.
[22:24] We also have security access right
[22:26] inside this and one of the coolest thing
[22:28] that you get to see which is what I mean
[22:29] was showing you that now you
[22:32] preemptively get all the compliance
[22:34] issues that are there in your agent. So,
[22:36] if you notice on the bottom, you'll see
[22:37] that there are a lot of failed
[22:39] issues that has happened. And this is
[22:41] again proactively
[22:44] agent platform is trying to figure out
[22:46] whether you know there has been issues
[22:48] or not.
[22:49] The last part is the evaluation which is
[22:51] part of the optimize. What you see here
[22:54] is different kinds of evaluation that
[22:56] your models go through and you can do
[22:59] couple of things. You can actually do
[23:00] online monitors which essentially means
[23:02] that you can define these logics and
[23:04] these logics will keep on monitoring
[23:06] each and every part of your agents
[23:07] through the span and the traces that I
[23:08] was showing you. And then you can
[23:10] actually see the status of them. You can
[23:12] see the traces where they have been
[23:13] triggered or you can sort of define your
[23:15] matrix. So, if you go to create metric,
[23:18] you now have a way where you can define
[23:19] metric either through prompt or through
[23:21] code. So, you can actually define a code
[23:23] and say that hey, this is what I want
[23:24] you to
[23:25] do this. So, this the part of
[23:28] evaluation. Now, one last thing that I
[23:29] want to show you is
[23:32] the Gemini Enterprise. So, Gemini
[23:33] Enterprise becomes the front of
[23:36] everything that you do, right? So, this
[23:38] is pretty much like saying you have
[23:39] Gemini as a front end of everything that
[23:41] you do in the back end. The way this
[23:43] looks is very interesting because now
[23:45] you can actually just
[23:48] Oops.
[23:52] So, this is how you see all the agents.
[23:55] So, for example, if I do at the rate,
[23:56] you can see my loan agent is there. And
[23:58] all your users can directly access these
[24:01] agents and these agents can then be
[24:03] follow all the policies or the
[24:04] guardrails and all of those things. And
[24:07] that's all. With that, I'm going to
[24:08] invite Thomas and Gautier to talk about
[24:11] how they are using agent platform to do
[24:12] some amazing stuff and scale things at
[24:14] L'Oreal. Thank
[24:25] Well, hello everyone. It's so good to be
[24:27] there
[24:28] for the third time in a row speaking
[24:30] about AI at L'Oreal.
[24:32] So, you have to know that the journey of
[24:34] L'Oreal have started in AI thanks to
[24:36] Google.
[24:38] But before jumping into what we do with
[24:39] Google, just a few words about who we
[24:42] are at L'Oreal for the one who don't
[24:44] know us.
[24:45] We are the world's largest beauty
[24:47] company in the world selling 7 billion
[24:50] units every year through 150 countries
[24:53] into with through 40 global brands.
[24:57] Also, we are the fourth largest
[25:00] advertiser worldwide.
[25:01] So, you can imagine that we are the
[25:03] leader in the world of beauty,
[25:05] but we are also the experts in the world
[25:07] of tech
[25:08] able to seize what starts.
[25:11] And to demonstrate you that, tell me let
[25:13] me tell you our story.
[25:16] Back in December '23,
[25:18] like many of you in this room, we took
[25:20] our first step into the GenAI era.
[25:23] We launched our L'Oreal GPT,
[25:26] a secure internal AI solution built for
[25:29] our employees.
[25:31] And honestly, we were proud.
[25:34] 2,000 weekly users felt like a real
[25:37] achievement at the beginning.
[25:39] But we moved fast and we quickly
[25:41] realized that keeping this solution
[25:43] siloed won't be enough
[25:46] if we truly wanted to scale AI across
[25:49] the company.
[25:50] We needed to open our own-made backend
[25:53] to the tech teams
[25:54] giving them the building blocks to
[25:56] develop their own use cases in a unified
[25:59] and scalable way.
[26:01] So, we transformed L'Oreal GPT into a
[26:03] set of reusable GenAI services.
[26:06] And that was our story of next '24.
[26:09] And those two small projects, they grew
[26:11] up very fast.
[26:13] And by September '25, we had already
[26:16] 30,000 weekly users on L'Oreal GPT
[26:19] and over 50 use cases in prod through
[26:21] the GenAI services
[26:23] sending approximately 15 million
[26:25] messages every month to the platform.
[26:29] And we started to have some zero code
[26:30] agents developed by the employees
[26:33] through L'Oreal GPT.
[26:35] We had built some We had We had built
[26:37] something real, something big.
[26:40] But with that scale came our truth.
[26:43] Continuing to build bulletproof
[26:45] enterprise-grade technology
[26:47] entirely at home was no longer the
[26:50] smartest path forward.
[26:52] So, we made a bold strategic decision
[26:55] from make
[26:56] to buy.
[26:57] We chose to leverage the best tech
[26:59] available, the Google Cloud Agent
[27:01] Platform that we just saw just before,
[27:03] to create the L'Oreal BeautyTech Agency
[27:06] Platform.
[27:07] Not because we couldn't build this,
[27:09] but because partnering with the
[27:11] best-in-class technology meant that we
[27:14] could focus on what truly matter for us,
[27:16] creating value for L'Oreal.
[27:19] And today, here at next '26,
[27:22] the results speak for themselves. 41,000
[27:25] weekly users on L'Oreal GPT
[27:27] and 30,000 zero code agents
[27:30] developed by the employees through this
[27:32] platform. All that running on ADK and
[27:35] Agent Engine.
[27:37] We're not just ready for the moon, we
[27:38] already have the rocket and we are in
[27:40] flight.
[27:42] So, what exactly do we do?
[27:45] At its core, the BeautyTech Agency
[27:47] Platform is simple.
[27:48] The single place where our L'Oreal tech
[27:50] teams makes and buys agents
[27:53] leveraging everything you saw from Aman
[27:56] and Levy.
[27:57] Think of it as a unified control tower.
[28:00] At the top, our everyday business
[28:02] applications.
[28:04] The one that we make,
[28:05] but also the one that we buy such as
[28:07] Salesforce, ServiceNow, SuccessFactors
[28:09] and so on.
[28:11] All connected and all benefiting from
[28:13] what sits underneath.
[28:15] The heart of the platform is the
[28:17] gateway.
[28:18] The layer that governs every model,
[28:21] every tool, every agent
[28:23] with built-in guardrails.
[28:25] Nothing goes through without control and
[28:28] human oversight.
[28:30] It's the place for the tech teams to
[28:32] train models, to build and deploy agents
[28:35] with ADK, Agent Engine and so on,
[28:38] to build tools with the MCP,
[28:40] and also to bring the knowledge that is
[28:44] mainly plugged on our data platform on
[28:46] BigQuery, but not only.
[28:49] And critically, this platform is open.
[28:51] External agents from our SaaS vendors,
[28:55] external models such as the Anthropic
[28:57] ones,
[28:58] external MCP servers,
[29:00] they are all plugged seamlessly
[29:02] by leveraging standard protocols such as
[29:05] A2A, MCP, UCP and so on.
[29:09] And we have also two final superpowers,
[29:11] the observability,
[29:13] so we always know what's happening,
[29:15] the discoverability, so every team can
[29:18] find and reuse what already exists.
[29:21] One platform, every agent, full control.
[29:25] But Gautier,
[29:27] can you tell us a bit more about what do
[29:29] we do to empower the team?
[29:31] Because you know, having the platform
[29:33] and the technology is not enough. And
[29:35] what do we need to bring that to life?
[29:38] Thank you, Gautier.
[29:39] Thank you, Thomas.
[29:41] So, yes, the Agent Platform is where we
[29:43] build agents, but building building
[29:45] agents is only part of the challenge.
[29:48] The real question is how do you move an
[29:51] an entire company into the agentic era?
[29:55] For that, we need five pillars. On the
[29:58] technical side, we have the agentic
[30:00] toolkit
[30:01] and the AI gateway.
[30:04] But the tech alone won't drive adoption.
[30:07] We also need a human stack with it.
[30:10] So we have a center of excellence,
[30:13] a governance task force, and a
[30:15] community. So
[30:17] let's start with the tech.
[30:20] The agentic toolkit is the bread and and
[30:24] butter of the platform. So developer
[30:27] must have everything they need to build,
[30:30] to scale, and to monitor their agents.
[30:34] We chose the GCP stack. Why?
[30:37] Because of everything we saw earlier,
[30:39] because it's the best stack.
[30:41] But here's the real magic.
[30:43] The platform evolves without us.
[30:47] With agentic demand expanding,
[30:50] we can't be on the critical path.
[30:53] So Google Google develops, improves,
[30:56] document services,
[30:58] while our team can focus on bringing
[31:01] value.
[31:03] They can have access instantly to the
[31:05] latest features.
[31:07] They can upskill,
[31:09] but they can also learn on industry
[31:11] standards.
[31:12] On our side, what is interesting is that
[31:15] we can hire people that are ready from
[31:18] day one.
[31:21] So now that the developer can build
[31:23] agents,
[31:25] we need to make sure we need we can
[31:27] handle all these.
[31:30] And how do we make
[31:31] how do we reuse components at scale?
[31:34] That's what the gateway is for.
[31:37] Think of it in three words:
[31:40] discover, reuse, govern.
[31:43] So discover.
[31:45] Models, MCPs, agents, guardrails, all
[31:48] registered in a single place.
[31:51] Every team can find what's already
[31:53] exists across the group.
[31:56] Then reuse.
[31:58] Based on standard protocols, we ensure
[32:01] we can reuse everything we build from
[32:03] all our business and technical
[32:05] platforms.
[32:07] And govern. We monitor usage in real
[32:10] time. We validate what gets deployed,
[32:14] and we drive
[32:15] economies of scales, especially in terms
[32:18] of model usage.
[32:21] So let's see at the typical project.
[32:25] Um so on the left, you have a growing
[32:28] ecosystem of agents, tools, and models
[32:32] made available through the gateway.
[32:35] As a use case, you pick the model you
[32:37] need, your MCP, your agents, and you
[32:41] wire them together with ADK as the
[32:43] backbone.
[32:45] Then you plug you plug the GCP services
[32:48] to handle sessions, memory, rag.
[32:51] You deploy with agent engine, and you're
[32:54] good to go.
[32:55] No proprietary framework to maintain, no
[32:58] waiting for the central team. You get
[33:01] access to the latest models, internal
[33:04] and external tools, and agents to build
[33:07] faster.
[33:09] It's that simple.
[33:11] Then, as I was saying, we need a human
[33:13] stack.
[33:15] So starting with a center of excellence,
[33:18] this is our team.
[33:19] So we are the platform experts providing
[33:22] global tools like starter packs, expert
[33:25] supports, and we also manage
[33:28] relationship with providers like
[33:30] Anthropic and Google.
[33:32] The key shift is that we went from
[33:34] building services for everyone
[33:37] to empowering everyone to build.
[33:39] That's a fundamental change
[33:42] in how we operate, and it's the only way
[33:44] to scale to the entire company.
[33:47] Then governance has to evolve, too.
[33:51] As agentic bring new complexities, and
[33:54] we move from overseeing the work of a
[33:56] few teams to ensuring that
[33:59] every AI builds across the company
[34:03] meets our standards in terms of
[34:06] enterprise architecture, cybersecurity,
[34:09] trustworthy AI,
[34:10] data privacy, and legal compliance.
[34:15] Okay. So we have the global team.
[34:18] And now how do we scale to all the 80
[34:20] teams across the company?
[34:22] That's where the community comes in.
[34:25] We created what we call the lineup.
[34:28] These are handpicked
[34:30] AI champr- champions across the
[34:33] organization.
[34:34] They are both technical experts that can
[34:37] tackle really complex use cases and
[34:40] business experts with deep knowledge of
[34:43] their business and organization.
[34:46] They are a a really relay network to
[34:49] make sure we cover all the company.
[34:52] And then learning.
[34:54] With AI, the the wall between IT and
[34:57] business is fading.
[35:01] It's now easier than ever to upskill, to
[35:03] learn advanced concepts, to bring value.
[35:07] And that is something we want to
[35:08] embrace.
[35:10] AI is not only for IT.
[35:12] We encourage everyone to learn and
[35:15] contribute.
[35:17] The goal, let no one behind.
[35:20] To make all of this accessible, we need
[35:22] to go beyond the GCP console.
[35:25] It's great for developers, but we need
[35:28] something for everyone.
[35:30] And that's why we why we are building
[35:32] the agentic platform web app.
[35:35] It will be the front door to the entire
[35:37] platform
[35:38] with a catalog to discover every
[35:41] components available in the gateway,
[35:43] easy access to training path,
[35:46] our internal processes, global
[35:49] monitoring dashboards, and a feed with
[35:51] the latest news from the agentic world.
[35:55] But here was here is what's the most
[35:58] exciting.
[35:59] We the we don't just build agents with
[36:03] the platform, but agents are at its
[36:06] core. The platform
[36:08] it runs on agents.
[36:10] And here's a first look of what it looks
[36:12] like.
[36:17] So in the platform, you have access to
[36:19] the catalog,
[36:21] the the
[36:22] principal items, but also resources that
[36:26] will come.
[36:27] You have access to the agents, both
[36:29] internal and external, that can be
[36:31] reused in your use case, and MCPs.
[36:35] Same thing.
[36:36] You can see the details of every single
[36:38] MCPs, the tools available, and how to
[36:41] connect them in your use case.
[36:45] And and you also have access to the
[36:47] models. Here, we really want to be able
[36:50] to compare, to really choose the right
[36:52] one for the usage.
[36:54] And finally, the guardrails.
[36:57] So these these are guardrails we want to
[36:59] share across multiple use cases.
[37:03] So
[37:04] uh just before finishing, I would like
[37:06] to give a big thank you to all the
[37:08] L'Oreal teams working on the agentic
[37:11] platform.
[37:12] And uh and uh yeah, big big kudos to the
[37:16] teams.
[37:17] Thank you.
[37:20] >> [applause]
[37:27] >> That was great. I that line that you had
[37:30] earlier really resonated with me from
[37:31] going from building services to enabling
[37:34] teams to be able to build products. That
[37:36] was really amazing. Well, if you're
[37:38] looking for what to do after this,
[37:40] there's a couple of quick QR codes you
[37:41] can scan. On the left is a resource for
[37:44] being able to check out the latest agent
[37:47] platform and get started. Takes you
[37:48] directly to our docs. And then there's
[37:50] also
[37:51] a uh sort of guidebook that we have on
[37:54] learning how agent platform can enable
[37:56] your team. So to bring it all together,
[37:59] this year with agent platform, you're
[38:00] getting a deeper ecosystem, investments
[38:03] in interoperability,
[38:05] ease of use so your entire team can work
[38:07] in one place, and customizability so
[38:10] that those agents actually work for your
[38:11] business. Build, scale, govern, optimize
[38:14] end to end in one platform. We can't
[38:16] wait to see what you all will build.
[38:18] Maybe come back again next year show us
[38:20] how things have changed again.
[38:24] >> [music]