# Great Agents Start with Great Infrastructure Akshat Bubna · AI Engineer NYC 2026 · October 13 [Watch recording](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16931s) [04:42:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16931s) Great! Sorry, I'm good. Hey everyone, I'm Akshat, CTO of Modal, and I'm going to talk [04:42:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16940s) about really need from their infrastructure. [04:42:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16945s) I'm told this audience is very finance adjacent, so I thought I'd talk about what [04:42:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16953s) if every cell in a spreadsheet could be backed by an agent? [04:42:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16959s) And imagine you're a financial analyst at a fund. [04:42:44](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16964s) you want to run a complex analysis on every company in your portfolio. [04:42:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16972s) Normally in the past, this would take days. [04:42:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16974s) You'd have to connect every portfolio company, [04:42:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16978s) look at their data sources, run some Python for each of them. [04:43:03](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16983s) now obviously with agents, this should be much easier. [04:43:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16986s) Well, actually, one of our customers, Hebbia, [04:43:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16989s) built something that's exactly this. [04:43:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16991s) you can go in right now in their product, [04:43:15](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16995s) you have a spreadsheet, [04:43:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=16998s) and you can run a query [04:43:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17000s) where literally every cell is [04:43:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17003s) backed by a different agent. [04:43:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17006s) And I wanted to talk about what does building this really look like. [04:43:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17014s) Well, first, every agent is backed by a modal sandbox. [04:43:41](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17021s) Sandbox is a secure environment for running untrusted code. [04:43:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17029s) It used to be that every sandbox was one-to-one with a container or micro VM. [04:43:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17036s) That's actually no longer the case. [04:43:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17038s) Today actually, modal has this feature where you can run multiple containers or micro VMs [04:44:05](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17045s) inside a sandbox. [04:44:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17046s) You actually need to do that for doing something like this, where in this architecture you [04:44:13](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17053s) run the harness. [04:44:15](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17055s) The harness is the agent loop that does things like running the, calling the model API for [04:44:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17061s) inference, running the tool calls, and then looping over it. [04:44:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17066s) want to separate that from the place where the tool calls actually run for security and [04:44:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17071s) isolation and a bunch of other things. [04:44:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17074s) So you have the sandbox that's running multiple containers, including a sidecar. [04:44:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17080s) You also want shared storage, because you want your agents to be able to talk to each [04:44:44](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17084s) other. [04:44:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17085s) And it's up to you how you stripe it. [04:44:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17087s) Like maybe you have every row gets the same volume, and agents in that row in the spreadsheet [04:44:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17094s) can collaborate. [04:44:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17096s) And you also want to share storage between your main container and the sidecar for some [04:45:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17101s) cases. [04:45:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17102s) And finally, you need all of this to spin up really, really fast, because your spreadsheet [04:45:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17107s) can have lots of cells. [04:45:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17108s) If someone wants to query, you want to launch thousands of agents, and then you want to [04:45:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17112s) spin them back down when it's over. [04:45:15](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17115s) So before I talk more about that, I just wanted to take a quick aside to talk about what is [04:45:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17122s) modal. [04:45:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17123s) Modal is a platform of primitives that is built for AI applications. [04:45:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17129s) We have built this platform from scratch, built our own file systems, scheduling and [04:45:35](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17135s) the runtime to give you great performance and devx for use cases like inference, including [04:45:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17142s) LLM inference, sandboxes, training and functions, running a lot of batch processing and what [04:45:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17148s) not. [04:45:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17149s) Today in particular I'm going to talk about sandboxes, what this primitive is and how [04:45:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17154s) it's evolving. [04:45:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17154s) So if you want to build something like this, or even further, and we'll talk more about [04:45:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17159s) what that means, what are the primitives you need for that? [04:46:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17164s) The first thing is, if you're building an agent, you need a workspace. [04:46:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17167s) And I have some examples from some of our current customers. [04:46:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17171s) You might have heard of Ramp Inspect. [04:46:13](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17173s) Ramp Inspect is Ramp's internal coding agent that I think at this point pushes almost every [04:46:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17181s) single PR at Ramp. [04:46:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17182s) If you're building a coding agent, you need your repo, your GitHub repo, your dependencies, [04:46:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17191s) you need your dev services, all in the same workspace for each agent. [04:46:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17198s) Another example is basis. [04:46:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17202s) it looks like is your documents might arrive, your agent is then triggered to work [04:46:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17208s) on them, but then it's waiting for a response from the accountant. [04:46:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17212s) You don't know how long this could take. [04:46:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17214s) It could take many days. [04:46:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17216s) When you want to persist the state over time, your usage is going to change. [04:47:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17222s) In the second chart, you can see usage for each sandbox. [04:47:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17227s) There's definitely some consistent trends, but it's all over the place. [04:47:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17231s) If you were provisioning this, it doesn't make sense. [04:47:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17234s) and then reliably get them when the sandbox needs them [04:47:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17237s) and only pay for them for that duration. [04:47:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17242s) The other interesting emerging thing is [04:47:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17245s) agents need a lot more cloud science [04:47:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17248s) where pre-processed some data on CPUs [04:47:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17252s) or spin-up GPUs for running pre-processed some data on CPUs [04:47:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17257s) or spin-up GPUs for running a scientific model, [04:47:41](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17261s) it will offload that to a modal sandbox. [04:47:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17263s) is a scientific model, [04:47:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17265s) it will offload that to a modal sandbox. [04:47:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17268s) And we're seeing this show up in RSI [04:47:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17270s) about what happens when you give an agent [04:47:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17273s) sandboxes to Hillclient. [04:47:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17277s) The next emerging thing is [04:47:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17279s) agents are collaborating a lot more, [04:48:03](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17283s) and that means that you share a file system [04:48:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17287s) with your sub-agent, [04:48:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17289s) you share file system. [04:48:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17291s) The next emerging thing is agents are collaborating a lot more, and that means that you share [04:48:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17300s) a file and files with your human where you're all collaborating with the same thing. [04:48:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17306s) This means file systems that are fast, have consistency, and devx you can reason about. [04:48:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17313s) Something we're working on with modal volumes. [04:48:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17316s) The related thing is we're seeing a lot more self-improvement. [04:48:41](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17321s) will save and update memory, they'll run background tasks, which is now called [04:48:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17327s) dreaming, to update this memory over time. [04:48:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17330s) Actually this also just looks like files in a file system. [04:48:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17334s) But again, you want really great performance from this file system. [04:49:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17342s) and the scary thing about agents is they can mess up and they don't really know when they might [04:49:11](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17351s) leak your sensitive data to a german forum or something so you need to have a lot of control [04:49:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17358s) over what they can and cannot do and you need things like controlling what ips can even talk [04:49:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17365s) to the agent, how do you authenticate it, what domains can it talk to, and also if you're [04:49:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17372s) using credentials for talking to services, basically you want to inject them in a way [04:49:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17377s) that the agent cannot actually see them and then exfiltrate them somewhere later. [04:49:44](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17384s) And the interesting evolution of this is we've realized that domain-related allow lists are [04:49:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17390s) actually not enough. [04:49:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17392s) a lot of our users actually want very specific control [04:49:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17396s) over even like what paths can I talk to [04:50:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17400s) on let's say Datadog or GitHub. [04:50:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17402s) Like maybe I want to reason about specific operations [04:50:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17406s) that the agent is doing. [04:50:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17407s) And going back to sidecars, [04:50:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17409s) we've seen this emerging as a pattern [04:50:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17410s) where users actually write their own lightweight proxy [04:50:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17414s) that has custom rules about what exactly you can do. [04:50:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17417s) And Ramp's another example of a company [04:50:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17421s) that's been using this successfully. [04:50:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17425s) And the final underrated thing is it's really important to have really great observability. [04:50:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17432s) Going back to the example of auto research and an agent doing a lot of training, what [04:50:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17437s) happens when the training run fails? [04:50:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17440s) You want your agent to be able to easily look at the logs of the failure, reason about what [04:50:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17445s) happened and do another iteration with a different set of parameters. [04:50:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17450s) While training is running, you want to look at the metrics, GPU utilization and make sure [04:50:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17457s) it's not stalled. [04:50:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17458s) Maybe you want to look at the budget and make sure you're not spending too much on your [04:51:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17462s) sandbox. [04:51:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17464s) And obviously all of this is really important for humans, but as agents are driving a lot [04:51:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17469s) more compute, we're spending a lot of time exposing all of this to the agents themselves [04:51:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17474s) so they can reason about all of this. [04:51:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17477s) All right. [04:51:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17478s) I think we know what we need. [04:51:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17480s) Why don't we have Astra build us an agent platform? [04:51:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17485s) Well, I don't think it's a good idea. [04:51:27](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17487s) I don't think you should vibe code a sandbox platform. [04:51:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17490s) And there's three main reasons for that. [04:51:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17492s) Number one is reliability. [04:51:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17494s) I think we're all painfully aware how hard it is to get right. [04:51:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17499s) Certain websites we use every day keep going down. [04:51:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17503s) and the reason for that is even with AI, [04:51:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17507s) it's actually really hard to reason about the feedback loop [04:51:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17510s) of what happens in the case of this tail event [04:51:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17513s) when the system breaks [04:51:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17514s) and this is why you need to be very careful in building systems. [04:51:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17519s) Security is another one. [04:52:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17520s) You could vibe code a sandbox platform [04:52:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17522s) but who is going to do all the work [04:52:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17524s) to continually reason about the security boundaries [04:52:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17527s) and make sure it's defensible [04:52:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17530s) in an age when everything is evolving very fast. [04:52:15](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17535s) And finally, scale. [04:52:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17537s) We have a blog post about how we scale to a million sandboxes, [04:52:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17541s) which involves a lot of very interesting engineering design [04:52:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17544s) and reasoning about systems. [04:52:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17546s) And this is, again, something that you kind of don't want to leave up [04:52:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17551s) to a Vibe-coded platform, [04:52:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17552s) because when you get to scale, you want it to work. [04:52:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17556s) And on that point, we're actually seeing massive scale [04:52:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17559s) for a lot of things I mentioned. [04:52:41](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17561s) So, for example, for the snapshots I mentioned, [04:52:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17565s) it's growing quite fast. [04:52:46](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17566s) We're seeing 20 million snapshots a week. [04:52:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17569s) We're processing about eight petabytes of data a day [04:52:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17574s) for each of these snapshots. [04:52:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17577s) And our users are creating almost 2.6 million volumes a week, [04:53:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17582s) and all of this is growing super fast. [04:53:05](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17585s) And just to take a customer workload for personal assistant, [04:53:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17590s) You need to create 10,000 sandboxes a second. [04:53:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17594s) You need to run almost a million sandboxes. [04:53:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17598s) And they're not going to last very long, they'll just last for like two or three minutes. [04:53:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17602s) But when they come up, [04:53:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17603s) you want them to be interactive in less than a second. That's what this workload really demands. [04:53:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17612s) And the other interesting thing on scale is you actually cannot get all of the scale from a single cloud platform. The world is actually getting crunched on CPUs as well, not just GPUs. [04:53:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17623s) So one thing we spend a lot of time doing is actually aggregating cloud capacity from [04:53:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17628s) a bunch of hyperscalers and neoclouds. [04:53:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17630s) So our users have access to CPU capacity across the world when they need it, but we cannot [04:53:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17636s) rely on one cloud provider to do that. [04:53:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17639s) Okay, so where is all this going? [04:54:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17642s) I hinted at this earlier, but the harnesses are really good at making use of dynamic environments. [04:54:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17650s) and what this is leading to is a lot of separation of the brains, [04:54:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17654s) which is the harness, from the hands, [04:54:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17657s) which is the sandbox or the runtime where stuff is running. [04:54:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17662s) And what this means, I have an example here, [04:54:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17665s) is the harnesses are going to be able to spin up environments [04:54:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17669s) based on what they need right now. [04:54:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17672s) They're good at snapshotting them and keeping them around [04:54:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17676s) as artifacts that they need later. [04:54:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17678s) And again, adjusting all of this as necessary. [04:54:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17682s) So I think this points at an interesting evolution [04:54:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17685s) of the pets to cattle analogy we're all familiar with. [04:54:51](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17691s) To recap, in the olden days, [04:54:55](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17695s) people thought of services like pets. [04:54:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17697s) Each one was bespoke. [04:54:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17699s) You had to manage it. [04:55:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17700s) You had to reason about this individual deployment specifically. [04:55:05](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17705s) And then over time, with containerization and Kubernetes, people started to think about [04:55:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17712s) it as homogenous templates. [04:55:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17716s) These servers are all the same, you can scale them up and down. [04:55:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17719s) If one of them fails, it's not an issue because our system can heal from that. [04:55:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17724s) And I think this is where a lot of people deploying agents are today. [04:55:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17729s) They're kind of using these homogenous environments and scaling up these agents. [04:55:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17732s) But I think that's really limiting because the models are so good at reasoning about [04:55:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17738s) dynamic environments, snapshotting, creating lots of environments when needed and working [04:55:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17743s) across them. [04:55:44](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17744s) So what does the new world look like? [04:55:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17748s) analogy here is it kind of looks like an ocean of microbes, of all these organisms [04:55:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17754s) that are self-evolving and working with each other. [04:55:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17758s) credits this fix for the analogy here. [04:56:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17761s) but one thing I'm really excited about [04:56:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17764s) is as you give agents these really rock-solid primitives [04:56:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17768s) for compute and storage and networking, [04:56:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17772s) they will be able to actually do a lot more [04:56:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17774s) by doing the stuff I talked about, [04:56:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17776s) like spinning up lots of environments, [04:56:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17778s) working with each other, and being very dynamic. [04:56:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17782s) So yeah, I and us at Modal really think [04:56:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17785s) that great primitives let your agents do more, [04:56:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17790s) and yeah if you're interested in talking about this more [04:56:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17792s) reach out to us [04:56:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=17793s) thank you