1 00:00:02,221 --> 00:00:08,041 Hi everyone, my name is Philipp and I work at [unclear] and happy second Gemini 4 day. 2 00:00:08,361 --> 00:00:10,761 So who of you already tried the new model? 3 00:00:11,021 --> 00:00:11,981 Can you raise your hands? 4 00:00:12,781 --> 00:00:14,381 Okay, that's not enough. 5 00:00:14,601 --> 00:00:17,981 So you should definitely take your time later, try it. 6 00:00:18,061 --> 00:00:18,841 It is really good. 7 00:00:19,041 --> 00:00:22,161 We are going to use it and show it in our talk as well. 8 00:00:22,341 --> 00:00:27,281 So we are going to look at how you can build and code your own autonomous Wall Street co-workers. 9 00:00:27,281 --> 00:00:32,141 so hopefully we have many finance people or engineers working in finance here. 10 00:00:32,341 --> 00:00:36,041 So quickly before we go into where we are coming from, 11 00:00:36,601 --> 00:00:38,861 two years ago we basically had a chat box. 12 00:00:39,001 --> 00:00:42,201 We asked some simple questions like what's the capital of France, 13 00:00:42,401 --> 00:00:44,261 who won the last European championship. 14 00:00:44,481 --> 00:00:48,421 Then we moved to assistance using tools on our own computer, 15 00:00:48,841 --> 00:00:51,661 having the first coding agents writing code for us. 16 00:00:51,801 --> 00:00:55,801 And now we are moving into this autonomous co-worker area 17 00:00:55,801 --> 00:00:59,841 where we basically give or create a digital employee 18 00:00:59,841 --> 00:01:02,101 where we give an agent its own computer, 19 00:01:02,361 --> 00:01:03,101 give it a task, 20 00:01:03,141 --> 00:01:06,561 and hopefully come back a few hours later with a good result. 21 00:01:06,921 --> 00:01:09,921 So what does a co-worker normally need? 22 00:01:10,001 --> 00:01:13,441 And I think that goes for AI and for us humans as well. 23 00:01:13,581 --> 00:01:16,461 We need a brain to be able to do our work. 24 00:01:16,481 --> 00:01:21,441 We need a computer where we can save artifacts we create, 25 00:01:21,661 --> 00:01:22,361 where we can iterate. 26 00:01:22,361 --> 00:01:24,841 We need some kind of a new playbook, right? 27 00:01:24,841 --> 00:01:30,261 when you hire someone new they are new to your environment to your domain you might have specific 28 00:01:30,261 --> 00:01:36,861 workflows so you need playbooks we need tools for them to be able to do that job they all need 29 00:01:36,861 --> 00:01:42,341 credentials in one or the other way then there might be some specific house rules on what you 30 00:01:42,341 --> 00:01:46,641 should do or what you should not do what you can do on your own or where you need to get some 31 00:01:46,641 --> 00:01:52,901 approvals and then of course there are hopefully some schedules for working hours when you can come 32 00:01:52,901 --> 00:01:58,861 in or when you have to work. And the same goes for our agents. And we are going to work 33 00:01:58,861 --> 00:02:03,281 through all of those seven steps and how you can build your own digital co-worker with 34 00:02:03,281 --> 00:02:08,041 Gemini in this talk. But before we go in, yesterday was the big day. We announced Gemini 35 00:02:08,041 --> 00:02:15,201 for Argon and made it available to you. Biggest new features, we have now a 1 million unified 36 00:02:15,201 --> 00:02:20,781 context window. This means Gemini can generate up to 1 million output tokens. And we see 37 00:02:20,781 --> 00:02:25,901 strong improvements especially on math and reasoning tasks where the model now has more than 64k 38 00:02:25,901 --> 00:02:32,481 token available to think to work through to reason through some proofs and for math for for 39 00:02:32,481 --> 00:02:38,861 math and then very also very good is like we tested that model heavily so we know there are 40 00:02:38,861 --> 00:02:45,401 always some thinking around hey gemini is just banks benchmarks but the model is really good 41 00:02:45,401 --> 00:02:46,861 you should definitely give it a try. 42 00:02:47,061 --> 00:02:49,161 We use it extensively at Google. 43 00:02:49,381 --> 00:02:52,321 There are already many projects by Google employees 44 00:02:52,321 --> 00:02:53,181 who have used it. 45 00:02:53,341 --> 00:02:56,321 The easiest way to get started is the Interactions API. 46 00:02:56,681 --> 00:02:58,161 It's literally four lines of code, 47 00:02:58,301 --> 00:03:01,261 which for you to take, you get your API key in AI Studio, 48 00:03:01,921 --> 00:03:04,081 aistudio.com slash API key, 49 00:03:04,161 --> 00:03:06,701 and then you can immediately send your request. 50 00:03:06,961 --> 00:03:08,861 It's also, I would say, quite affordable 51 00:03:08,861 --> 00:03:12,481 for a frontier model with $2 in and $10 out. 52 00:03:12,521 --> 00:03:15,261 You should definitely try if you have a free minute later 53 00:03:15,261 --> 00:03:18,161 or even if you have some agents running right now, 54 00:03:18,321 --> 00:03:20,981 give it a link, tell it, hey, please set up Gemini 4 55 00:03:20,981 --> 00:03:21,761 and give it a try. 56 00:03:22,101 --> 00:03:24,581 So what have Googlers done so far with the model? 57 00:03:24,741 --> 00:03:27,501 I think it's very impressive because it's like a super new model. 58 00:03:27,641 --> 00:03:30,541 We had it available internally to test for a few weeks already, 59 00:03:30,561 --> 00:03:33,061 but Googlers are using it so heavily 60 00:03:33,061 --> 00:03:34,941 and shipping so many cool things. 61 00:03:34,961 --> 00:03:37,701 We had the Chrome team basically accelerating 62 00:03:37,701 --> 00:03:39,921 some decoupling of the Chromium browser. 63 00:03:40,081 --> 00:03:42,721 We have Gemma, which is our open model 64 00:03:42,721 --> 00:03:45,701 where Gemini accelerated Gemma 65 00:03:45,701 --> 00:03:47,901 to be more easily and more efficient to run. 66 00:03:48,081 --> 00:03:51,541 The ADK team from cloud improved their framework 67 00:03:51,541 --> 00:03:53,381 and then many, many other things, 68 00:03:53,561 --> 00:03:55,941 and especially the right one at the bottom. 69 00:03:55,941 --> 00:03:58,201 It's like we use it internally heavily 70 00:03:58,201 --> 00:03:59,701 and we see strong improvements 71 00:03:59,701 --> 00:04:02,701 for our internal code base as well. 72 00:04:03,001 --> 00:04:04,641 So what have people built? 73 00:04:04,781 --> 00:04:05,781 There are some demos. 74 00:04:06,141 --> 00:04:06,941 You can look them up, 75 00:04:07,121 --> 00:04:09,381 but there's one specific one from a colleague of mine 76 00:04:09,381 --> 00:04:10,241 which I want to show. 77 00:04:10,241 --> 00:04:17,101 he basically prompted Gemini to build a custom 3D ray tracer in Python. 78 00:04:17,321 --> 00:04:23,021 And it was completely built from Gemini and nothing external tools and at the end just Python. 79 00:04:23,261 --> 00:04:29,541 And it also used a Gemini TTS model and Luria to add some voiceover to explain how it works. 80 00:04:29,681 --> 00:04:33,521 And it's like very impressive for me how far we can push those models now. 81 00:04:33,721 --> 00:04:35,421 So try to be very ambitious. 82 00:04:35,721 --> 00:04:38,661 So how can we build our agent? 83 00:04:38,661 --> 00:04:42,781 And in Gemini API, we have a product called Managed Agents. 84 00:04:42,941 --> 00:04:47,781 And Managed Agents is basically a simple way for you to build a remote agent 85 00:04:47,781 --> 00:04:52,981 where you give Gemini a sandbox which it can use to run tools. 86 00:04:53,121 --> 00:04:58,001 So when we normally build an agent, we have this weird tool calling group, right? 87 00:04:58,061 --> 00:04:59,641 We send a request to the model. 88 00:04:59,781 --> 00:05:02,761 The model might generate a JSON function call. 89 00:05:02,901 --> 00:05:04,481 We receive that function call. 90 00:05:04,621 --> 00:05:07,581 We run that function call on our machine, execute the code, 91 00:05:07,581 --> 00:05:09,081 and then we go back to the model. 92 00:05:09,301 --> 00:05:12,001 And with managed agents, we basically provide a sandbox 93 00:05:12,001 --> 00:05:14,901 which Gemini can directly use to run all of those calls. 94 00:05:15,081 --> 00:05:18,141 So if it decides to execute code to write a file, 95 00:05:18,321 --> 00:05:21,341 it has its own computer available for it to use, 96 00:05:21,361 --> 00:05:23,561 and it takes just an API call. 97 00:05:23,741 --> 00:05:27,341 So instead of providing the model ID with Gemini 4 Argon, 98 00:05:27,461 --> 00:05:29,401 you provide the agent ID. 99 00:05:29,541 --> 00:05:32,221 In this case, it's anti-gravity preview 09. 100 00:05:32,881 --> 00:05:33,561 Give it a prompt. 101 00:05:33,901 --> 00:05:35,081 The environment is remote, 102 00:05:35,101 --> 00:05:37,221 so you want to run it in the managed sandbox. 103 00:05:37,581 --> 00:05:38,741 and that's all it takes. 104 00:05:39,041 --> 00:05:41,861 So let's get into building Buck. 105 00:05:42,081 --> 00:05:45,601 Buck is our co-worker for today, and Buck needs a brain. 106 00:05:45,901 --> 00:05:48,501 Our brain is going to be Gemini, 107 00:05:48,741 --> 00:05:50,381 our computer is going to be Agents, 108 00:05:50,461 --> 00:05:52,121 and then we all love all of the other things. 109 00:05:52,261 --> 00:05:54,141 So how can you create an agent? 110 00:05:54,601 --> 00:05:57,261 In Managed Agents, you have a very simple API 111 00:05:57,261 --> 00:05:59,321 where you can define your own agent. 112 00:05:59,501 --> 00:06:00,321 So you give it an ID. 113 00:06:00,601 --> 00:06:02,601 That's the name which you are going to use later 114 00:06:02,601 --> 00:06:03,481 to call the agent. 115 00:06:03,881 --> 00:06:07,181 We have the base agent, so that's the main agent underlying. 116 00:06:07,581 --> 00:06:09,481 In this case, it's the agent harness. 117 00:06:09,741 --> 00:06:14,441 So the remote agent uses the same agent harness, 118 00:06:14,601 --> 00:06:16,481 which we have in Antigravity as well. 119 00:06:16,621 --> 00:06:19,481 So if you are using Antigravity as your coding agent, 120 00:06:19,641 --> 00:06:22,261 you get the same agent harness remotely. 121 00:06:22,521 --> 00:06:24,181 Then we want to use the new model. 122 00:06:24,201 --> 00:06:25,541 So we define our model. 123 00:06:25,741 --> 00:06:27,961 We have a system instruction, you are back. 124 00:06:28,381 --> 00:06:33,121 And then all it takes to call it, again, interactions API, create. 125 00:06:33,321 --> 00:06:36,381 And instead of using the model ID for Gemini 126 00:06:36,381 --> 00:06:39,401 or the default agent, we use our own agent buck. 127 00:06:40,141 --> 00:06:42,481 And our agent needs a computer, 128 00:06:42,661 --> 00:06:45,681 and the computer normally needs some additional files, 129 00:06:45,821 --> 00:06:46,961 additional scaffolding. 130 00:06:47,141 --> 00:06:50,021 So with managed agents, you can provide your skills 131 00:06:50,021 --> 00:06:52,721 or an agents MD file, maybe from your GitHub repository. 132 00:06:52,941 --> 00:06:55,281 You can inline those files directly. 133 00:06:55,281 --> 00:06:57,281 You can provide a GitHub repository, 134 00:06:57,621 --> 00:06:59,841 private or public, with your skills, 135 00:06:59,861 --> 00:07:02,061 and you can connect GCS buckets. 136 00:07:02,061 --> 00:07:04,081 So you can already bootstrap the environment 137 00:07:04,081 --> 00:07:05,701 the agent is going to work in. 138 00:07:05,701 --> 00:07:09,761 and everything in managed agent lives as a file. 139 00:07:09,961 --> 00:07:13,781 So we have skills, we have hooks, we have system instructions. 140 00:07:13,781 --> 00:07:17,561 At the end, it's just a file system or the agent's own computer 141 00:07:17,561 --> 00:07:21,421 where it can create new skills if it thinks it needs to improve itself, 142 00:07:21,621 --> 00:07:24,901 where it can store memories from previous sessions or learning. 143 00:07:25,181 --> 00:07:27,401 So at the end, we just work with files. 144 00:07:27,901 --> 00:07:31,741 And for knowledge or for our agent, 145 00:07:31,781 --> 00:07:34,921 I like to separate those in three different categories. 146 00:07:34,921 --> 00:07:39,821 So we have knowledge which we always need to include into every request. 147 00:07:40,021 --> 00:07:41,761 That's normally the system instructions. 148 00:07:41,761 --> 00:07:43,221 We want to be very concise. 149 00:07:43,421 --> 00:07:48,761 We want to try to keep it very short because that's something we need to pay on every request. 150 00:07:48,861 --> 00:07:53,801 So if we have very long system instructions, all of our requests are going to be much more expensive. 151 00:07:53,801 --> 00:08:01,621 If we have concise, then we have knowledge which the agent can decide when it needs to use them or not. 152 00:08:01,701 --> 00:08:03,601 So that's typically with skills, right? 153 00:08:03,601 --> 00:08:09,221 with skills we have very short descriptions which give the agent some guidance on like hey you can 154 00:08:09,221 --> 00:08:13,861 use that skill if the user wants to do a certain task or if you have to do certain tasks and then 155 00:08:13,861 --> 00:08:14,001 can 156 00:08:14,021 --> 00:08:15,981 the agent can read that skill. 157 00:08:16,201 --> 00:08:24,081 But for skills, we also want to make sure that we are not having super long, dense skills because that's also a high cost at the end. 158 00:08:24,241 --> 00:08:34,121 Context is very valuable for us, so we want to decouple them and separate them into much more slices where we can then have dedicated references and scripts 159 00:08:34,121 --> 00:08:38,161 if we want the agent to do very certain things 160 00:08:38,161 --> 00:08:39,481 where it needs more context. 161 00:08:40,021 --> 00:08:42,121 Okay, so let's look at an example skill 162 00:08:42,121 --> 00:08:43,181 which we are going to use. 163 00:08:43,301 --> 00:08:44,701 So we have a portfolio audit. 164 00:08:44,901 --> 00:08:46,081 The agent should use it. 165 00:08:46,181 --> 00:08:48,041 It has some instructions available for it, 166 00:08:48,061 --> 00:08:48,741 how to run it. 167 00:08:48,821 --> 00:08:49,981 So we have a Python script. 168 00:08:50,281 --> 00:08:51,241 The agent can run, 169 00:08:51,301 --> 00:08:52,701 and then some additional information 170 00:08:52,701 --> 00:08:54,441 on what it needs to do with the results 171 00:08:54,441 --> 00:08:55,401 and where to store it. 172 00:08:57,581 --> 00:09:00,581 Our fourth part of our onboarding 173 00:09:00,581 --> 00:09:02,081 for our coworkers tools, right? 174 00:09:02,081 --> 00:09:09,521 We humans use normal software for agents that's either built-in skills like Google Search, code executions, or MCP server. 175 00:09:09,721 --> 00:09:15,181 Or if you don't have an MCP server and rather want to write a local function, that's also possible. 176 00:09:15,381 --> 00:09:18,301 So, managed agents, we still have our agents create. 177 00:09:18,401 --> 00:09:22,421 We are now extending it to define our tools. 178 00:09:22,681 --> 00:09:26,221 We have built-in tools from the Gemini API for Google Search. 179 00:09:26,421 --> 00:09:30,001 So, there's nothing you need to additionally do to get Google Search integrated. 180 00:09:30,001 --> 00:09:32,821 but if you want to use, for example, spreadsheets, 181 00:09:33,501 --> 00:09:37,701 Workspace has now an MCP server available for Gmail for spreadsheets, 182 00:09:37,781 --> 00:09:38,601 so we can use that. 183 00:09:39,141 --> 00:09:43,181 Or if you want to provide your own functions or MCP server, 184 00:09:43,301 --> 00:09:44,141 that's also possible. 185 00:09:44,601 --> 00:09:47,221 So for your own function, it still works the same way 186 00:09:47,221 --> 00:09:48,941 as you normally do with models. 187 00:09:49,081 --> 00:09:50,761 So the agent creates a function call, 188 00:09:50,801 --> 00:09:53,021 and then you send back the function result. 189 00:09:53,301 --> 00:09:57,501 So it's basically for you a way to have agentic capabilities 190 00:09:57,501 --> 00:10:00,741 outside of these agents' environment and computers. 191 00:10:00,901 --> 00:10:02,861 So if you want to do certain things 192 00:10:02,861 --> 00:10:05,621 where the agent should not run the execution in the sandbox, 193 00:10:05,981 --> 00:10:07,341 that's a way for you to do it. 194 00:10:07,721 --> 00:10:09,781 And then we are logins. 195 00:10:09,861 --> 00:10:11,821 And I think that's, for me, the most exciting 196 00:10:11,821 --> 00:10:14,161 and the greatest feature about the Managed Agents API 197 00:10:14,161 --> 00:10:16,381 because credentials are super hard 198 00:10:16,381 --> 00:10:18,401 and super easy to get wrong with agents. 199 00:10:18,581 --> 00:10:21,021 And I think the previous talk already also mentioned 200 00:10:21,021 --> 00:10:24,081 that we are very careful about using agent 201 00:10:24,081 --> 00:10:26,201 because things can go wrong very easily. 202 00:10:26,201 --> 00:10:29,201 And we've thought a lot about this with managed agents, 203 00:10:29,241 --> 00:10:31,961 and we have support for three different ways of credentials, 204 00:10:31,961 --> 00:10:34,621 which makes it secure for your agent to use them 205 00:10:34,621 --> 00:10:36,681 and provide an isolated way. 206 00:10:36,721 --> 00:10:39,941 So we have bearer tokens for normal API calls, 207 00:10:40,121 --> 00:10:42,161 very familiar if you are doing software engineering. 208 00:10:42,161 --> 00:10:46,721 We have OAuth for all of the consumer-oriented MCP servers, 209 00:10:46,721 --> 00:10:47,421 all like tools. 210 00:10:47,581 --> 00:10:49,601 And then, of course, we need environment variables 211 00:10:49,601 --> 00:10:52,561 if we want to use certain CLIs or other features. 212 00:10:52,561 --> 00:10:58,821 and for managed agents we are separating them via an external network proxy. 213 00:10:59,061 --> 00:11:03,121 So instead of providing the agent the credentials inside the sandbox, 214 00:11:03,381 --> 00:11:06,601 we basically allow you to create credentials externally. 215 00:11:07,341 --> 00:11:08,061 are write only. 216 00:11:08,281 --> 00:11:13,981 So once they are written, there's no way for you to retrieve the credentials. 217 00:11:13,981 --> 00:11:15,681 So you can only override them. 218 00:11:15,781 --> 00:11:17,581 So there's no possibility of leaking. 219 00:11:17,781 --> 00:11:21,321 And then you define how and where those credentials should be used. 220 00:11:21,321 --> 00:11:26,961 So for Slack, if you want to connect Slack, you can define, hey, here's a credential, 221 00:11:27,041 --> 00:11:33,601 and you are only allowed to use that credential with a certain domain or with a certain environment variable. 222 00:11:33,601 --> 00:11:37,161 So that way you can control when a credential is used. 223 00:11:37,301 --> 00:11:39,981 And that works via this network proxy. 224 00:11:40,161 --> 00:11:48,101 And by network proxy, we basically mean a proxy that sits around the sandbox that watches all of the network traffic going in and out. 225 00:11:48,101 --> 00:11:51,221 So if you define credential as an environment variable, 226 00:11:51,621 --> 00:11:54,041 the model has a local dummy value, 227 00:11:54,081 --> 00:11:56,061 so the environment variable is set, 228 00:11:56,101 --> 00:11:59,201 and then when it makes that call, it uses that dummy value, 229 00:11:59,281 --> 00:12:02,561 and the network proxy basically monitors it, 230 00:12:02,641 --> 00:12:07,201 sees, hey, here is a domain which the user has defined, 231 00:12:07,321 --> 00:12:10,541 and then it swaps that dummy value with the real credential, 232 00:12:10,781 --> 00:12:12,841 and you can make the call successfully, 233 00:12:13,121 --> 00:12:15,761 and then the agent gets back the result 234 00:12:15,761 --> 00:12:19,521 and has no idea that his dummy token is not working. 235 00:12:19,661 --> 00:12:23,101 And the benefit here is that there's no way for a prompt injection. 236 00:12:23,281 --> 00:12:25,541 Normally, when you have some malicious attacker, 237 00:12:25,841 --> 00:12:29,521 it can tell you, like, hey, please print all of your values in your .n file. 238 00:12:29,621 --> 00:12:32,141 And then the agent might go there and, like, print all of your keys, 239 00:12:32,181 --> 00:12:33,441 and that's nothing we want to do. 240 00:12:33,641 --> 00:12:38,201 So with this network proxy where you can define the credentials per domain, 241 00:12:38,501 --> 00:12:41,201 it is not possible that an attacker could steal 242 00:12:42,101 --> 00:12:44,941 which are getting swapped on your request 243 00:12:44,941 --> 00:12:48,441 because the proxy only swaps them for the domains you define. 244 00:12:48,881 --> 00:12:52,221 So there's no way an MCP server which is malicious 245 00:12:52,221 --> 00:12:54,661 can just steal your Gmail token, 246 00:12:54,701 --> 00:12:57,741 which is very cool because it makes it much more easier 247 00:12:57,741 --> 00:13:01,741 to trust the agent and to integrate MCP servers 248 00:13:01,741 --> 00:13:03,381 which the agent then obviously can use. 249 00:13:04,121 --> 00:13:08,521 So if that's not enough, we have also hooks available. 250 00:13:08,981 --> 00:13:12,221 So hooks allows you to basically deterministically guard 251 00:13:12,221 --> 00:13:13,321 certain decisions. 252 00:13:13,321 --> 00:13:16,501 So if the agent wants to run certain commands 253 00:13:16,501 --> 00:13:18,061 which you might want to monitor 254 00:13:18,061 --> 00:13:20,581 or where you need some human in the loom approval, 255 00:13:20,821 --> 00:13:23,521 which is not part of the normal agentic conversation, 256 00:13:23,801 --> 00:13:25,121 you can define your hooks, 257 00:13:25,141 --> 00:13:27,761 very similar to all of the other agent harnesses 258 00:13:27,761 --> 00:13:28,721 where you have a matcher, 259 00:13:28,841 --> 00:13:30,301 which is basically regex, 260 00:13:30,361 --> 00:13:32,801 and then every tool call which the agent generates 261 00:13:32,801 --> 00:13:34,001 checks if there's a match, 262 00:13:34,181 --> 00:13:35,621 and then writes those hooks. 263 00:13:35,781 --> 00:13:37,541 And those hooks are also files, 264 00:13:37,661 --> 00:13:38,561 but they are protected, 265 00:13:38,561 --> 00:13:40,661 so the agent cannot just go there 266 00:13:40,661 --> 00:13:42,061 and changes a hook 267 00:13:42,061 --> 00:13:46,021 and basically tries to find a way around it. 268 00:13:46,881 --> 00:13:49,341 Okay, so last but not least, 269 00:13:49,521 --> 00:13:53,521 we need a way to get our agent more proactive. 270 00:13:53,781 --> 00:13:55,761 And currently we support one kind of triggers, 271 00:13:56,661 --> 00:13:58,901 basic cron triggers, which you can set up, 272 00:13:59,001 --> 00:14:03,281 schedule every 30 days a month at a.m. 273 00:14:03,321 --> 00:14:05,501 You want the agent to run a certain trigger 274 00:14:05,501 --> 00:14:08,181 with a certain input in the same environment 275 00:14:08,181 --> 00:14:10,721 to do maybe some due diligence, some auditing. 276 00:14:10,721 --> 00:14:13,141 and we are also working on webhooks 277 00:14:13,141 --> 00:14:15,401 that you can even make this much more interactive 278 00:14:15,401 --> 00:14:18,321 that you could connect external webhooks 279 00:14:18,321 --> 00:14:20,601 like from some certain SAC provider 280 00:14:20,601 --> 00:14:23,541 or maybe other providers you work together with 281 00:14:23,541 --> 00:14:25,521 which would directly then call your agent. 282 00:14:25,681 --> 00:14:27,301 The agent does some first auditing 283 00:14:27,301 --> 00:14:29,801 and then basically can call your Slack 284 00:14:29,801 --> 00:14:31,941 or your Microsoft Teams directly 285 00:14:31,941 --> 00:14:33,741 without you even needing to manage 286 00:14:33,741 --> 00:14:35,421 some infrastructure at all. 287 00:14:37,501 --> 00:14:39,041 So that's that. 288 00:14:39,041 --> 00:14:41,901 final picture and now let's look 289 00:14:41,901 --> 00:14:43,821 at some real demos 290 00:14:43,821 --> 00:14:45,881 and let's hopefully hope that 291 00:14:45,881 --> 00:14:47,161 the wifi is good enough. 292 00:14:48,361 --> 00:14:50,081 So I told Gemini 293 00:14:50,081 --> 00:14:52,121 that I want to do a cool Google Sheets demo 294 00:14:52,121 --> 00:14:54,121 and it helped me create basically this whole 295 00:14:54,121 --> 00:14:55,601 Google Sheets replication in 296 00:14:55,601 --> 00:14:57,541 my browser and 297 00:14:57,541 --> 00:15:00,061 when I click here we basically 298 00:15:00,061 --> 00:15:02,181 simulate this cron trigger so we send 299 00:15:02,181 --> 00:15:04,181 a real API call to managed 300 00:15:04,181 --> 00:15:06,081 agents. It should use our 301 00:15:06,081 --> 00:15:07,541 portfolio auditing skill 302 00:15:07,541 --> 00:15:09,681 and it runs the request. 303 00:15:09,901 --> 00:15:11,541 So first step basically is like, 304 00:15:11,561 --> 00:15:13,941 hey, I need to read the skill. 305 00:15:14,521 --> 00:15:16,021 It gets the result back. 306 00:15:16,021 --> 00:15:18,401 So it tries to understand what it should do here. 307 00:15:18,501 --> 00:15:20,701 It's like, I hope that's roughly accurate 308 00:15:20,701 --> 00:15:22,241 on like the Google Sheets. 309 00:15:22,781 --> 00:15:25,381 But it basically runs through all of those calls, 310 00:15:25,441 --> 00:15:26,901 calls financial services, 311 00:15:27,161 --> 00:15:27,981 tries to see, okay, 312 00:15:28,801 --> 00:15:30,761 did the market or the stock price change? 313 00:15:30,921 --> 00:15:31,681 And then at the end, 314 00:15:31,921 --> 00:15:34,781 hopefully it should update our own Google Sheets 315 00:15:34,781 --> 00:15:36,441 completely automatically. 316 00:15:36,441 --> 00:15:40,421 so it's still working through it, it has some first result 317 00:15:41,481 --> 00:15:44,441 and it updates the Google Sheet based on the trigger 318 00:15:44,441 --> 00:15:48,421 and that trigger would run every 8.30am without you needing to do 319 00:15:48,421 --> 00:15:51,701 anything and then of course if the agent has access to Slack 320 00:15:52,481 --> 00:15:56,361 then it can even post in some specific trading channel 321 00:15:56,361 --> 00:16:00,141 okay what changed, what is going on, how does it work and 322 00:16:00,141 --> 00:16:04,141 here we have an example where I can just 323 00:16:04,141 --> 00:16:05,181 ask Buck 324 00:16:06,981 --> 00:16:07,441 any 325 00:16:08,121 --> 00:16:08,581 recommendation 326 00:16:12,141 --> 00:16:14,041 and hopefully this works as well 327 00:16:14,041 --> 00:16:15,981 so it takes a little bit of time, you have seen in the 328 00:16:15,981 --> 00:16:17,741 other demo, we have seen all of that 329 00:16:17,741 --> 00:16:19,901 perfect, so there are 330 00:16:19,901 --> 00:16:21,801 three positions, so that's based on like this 331 00:16:22,741 --> 00:16:23,661 thread content 332 00:16:23,661 --> 00:16:26,061 so it has all of the previous turns available 333 00:16:26,061 --> 00:16:27,921 and it provides the context of 334 00:16:27,921 --> 00:16:29,981 what changed, but it's still basically 335 00:16:29,981 --> 00:16:31,941 not taking actions based on 336 00:16:31,941 --> 00:16:33,961 the system instructions you gave and all of 337 00:16:33,961 --> 00:16:37,461 the infrastructure you need to run that agent is 338 00:16:37,461 --> 00:16:41,901 part of the managed agents environment. So it's very easy for you to get started, to 339 00:16:41,901 --> 00:16:45,641 integrate those into the tools you already use, provide 340 00:16:45,641 --> 00:16:48,681 the systems you need, and then to get started. 341 00:16:53,901 --> 00:16:55,941 Back to the slide. So the whole 342 00:16:55,941 --> 00:16:59,641 co-worker, which we need in one service, we have the brain, Gemini. 343 00:16:59,641 --> 00:17:01,901 We have the computer with managed agents. 344 00:17:01,901 --> 00:17:05,361 We can provide skills for the same agents you use, 345 00:17:05,361 --> 00:17:06,901 maybe on your local computer. 346 00:17:07,121 --> 00:17:08,221 You can use the tools. 347 00:17:08,341 --> 00:17:11,301 You can provide MCP server for external services, 348 00:17:11,481 --> 00:17:12,661 data providers to connect. 349 00:17:12,861 --> 00:17:15,421 We have a secure way for you to provide credentials 350 00:17:15,421 --> 00:17:16,621 to rotate credentials. 351 00:17:17,101 --> 00:17:20,121 We have hooks to protect certain deterministic logic. 352 00:17:20,361 --> 00:17:21,561 And of course, we have triggers 353 00:17:21,561 --> 00:17:23,981 to much more proactively integrate it into it. 354 00:17:24,381 --> 00:17:26,001 So thanks for listening, 355 00:17:26,061 --> 00:17:28,381 and hopefully go hire your first AI co-worker. 356 00:17:29,641 --> 00:17:30,701 Thank you.