# Great Agents Start with Great Infrastructure Akshat Bubna · AI Engineer NYC 2026 · October 13 [Watch recording](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18528s) [05:08:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18528s) Great! Sorry, I'm good. Hey everyone, I'm Akshat, CTO of Modal, and I'm going to talk [05:08:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18536s) about really need from their infrastructure. [05:09:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18541s) I'm told this audience is very finance adjacent, so I thought I'd talk about what [05:09:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18550s) if every cell in a spreadsheet could be backed by an agent? [05:09:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18556s) And imagine you're a financial analyst at a fund. [05:09:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18560s) you want to run a complex analysis on every company in your portfolio. [05:09:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18568s) Normally in the past, this would take days. [05:09:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18571s) You'd have to connect every portfolio company, [05:09:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18574s) look at their data sources, run some Python for each of them. [05:09:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18579s) now obviously with agents, this should be much easier. [05:09:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18583s) Well, actually, one of our customers, Hebbia, [05:09:46](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18586s) built something that's exactly this. [05:09:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18588s) you can go in right now in their product, [05:09:51](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18591s) you have a spreadsheet, [05:09:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18594s) and you can run a query [05:09:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18597s) where literally every cell is [05:10:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18600s) backed by a different agent. [05:10:03](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18603s) And I wanted to talk about what does building this really look like. [05:10:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18610s) Well, first, every agent is backed by a modal sandbox. [05:10:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18617s) Sandbox is a secure environment for running untrusted code. [05:10:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18625s) It used to be that every sandbox was one-to-one with a container or micro VM. [05:10:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18632s) That's actually no longer the case. [05:10:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18634s) Today actually, modal has this feature where you can run multiple containers or micro VMs [05:10:41](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18641s) inside a sandbox. [05:10:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18642s) You actually need to do that for doing something like this, where in this architecture you [05:10:49](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18649s) run the harness. [05:10:51](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18651s) The harness is the agent loop that does things like running the, calling the model API for [05:10:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18657s) inference, running the tool calls, and then looping over it. [05:11:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18662s) want to separate that from the place where the tool calls actually run for security and [05:11:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18668s) isolation and a bunch of other things. [05:11:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18670s) So you have the sandbox that's running multiple containers, including a sidecar. [05:11:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18676s) You also want shared storage, because you want your agents to be able to talk to each [05:11:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18681s) other. [05:11:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18681s) And it's up to you how you stripe it. [05:11:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18684s) Like maybe you have every row gets the same volume, and agents in that row in the spreadsheet [05:11:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18690s) can collaborate. [05:11:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18692s) And you also want to share storage between your main container and the sidecar for some [05:11:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18697s) cases. [05:11:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18698s) And finally, you need all of this to spin up really, really fast, because your spreadsheet [05:11:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18703s) can have lots of cells. [05:11:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18705s) If someone wants to query, you want to launch thousands of agents, and then you want to [05:11:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18708s) spin them back down when it's over. [05:11:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18712s) So before I talk more about that, I just wanted to take a quick aside to talk about what is [05:11:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18718s) modal. [05:11:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18719s) Modal is a platform of primitives that is built for AI applications. [05:12:05](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18725s) We have built this platform from scratch, built our own file systems, scheduling and [05:12:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18732s) the runtime to give you great performance and devx for use cases like inference, including [05:12:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18738s) LLM inference, sandboxes, training and functions, running a lot of batch processing and what [05:12:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18745s) not. [05:12:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18746s) Today in particular I'm going to talk about sandboxes, what this primitive is and how [05:12:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18750s) it's evolving. [05:12:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18751s) So if you want to build something like this, or even further, and we'll talk more about [05:12:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18756s) what that means, what are the primitives you need for that? [05:12:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18760s) The first thing is, if you're building an agent, you need a workspace. [05:12:44](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18764s) And I have some examples from some of our current customers. [05:12:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18768s) You might have heard of Ramp Inspect. [05:12:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18770s) Ramp Inspect is Ramp's internal coding agent that I think at this point pushes almost every [05:12:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18777s) single PR at Ramp. [05:12:59](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18779s) If you're building a coding agent, you need your repo, your GitHub repo, your dependencies, [05:13:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18787s) you need your dev services, all in the same workspace for each agent. [05:13:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18794s) Another example is basis. [05:13:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18798s) it looks like is your documents might arrive, your agent is then triggered to work [05:13:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18804s) on them, but then it's waiting for a response from the accountant. [05:13:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18808s) You don't know how long this could take. [05:13:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18810s) It could take many days. [05:13:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18812s) When you want to persist the state over time, your usage is going to change. [05:13:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18818s) In the second chart, you can see usage for each sandbox. [05:13:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18823s) There's definitely some consistent trends, but it's all over the place. [05:13:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18827s) If you were provisioning this, it doesn't make sense. [05:13:51](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18831s) and then reliably get them when the sandbox needs them [05:13:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18834s) and only pay for them for that duration. [05:13:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18838s) The other interesting emerging thing is [05:14:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18841s) agents need a lot more cloud science [05:14:04](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18844s) where pre-processed some data on CPUs [05:14:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18848s) or spin-up GPUs for running pre-processed some data on CPUs [05:14:13](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18853s) or spin-up GPUs for running a scientific model, [05:14:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18857s) it will offload that to a modal sandbox. [05:14:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18859s) is a scientific model, [05:14:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18862s) it will offload that to a modal sandbox. [05:14:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18864s) And we're seeing this show up in RSI [05:14:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18866s) about what happens when you give an agent [05:14:29](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18869s) sandboxes to Hillclient. [05:14:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18874s) The next emerging thing is [05:14:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18876s) agents are collaborating a lot more, [05:14:39](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18879s) and that means that you share a file system [05:14:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18883s) with your sub-agent, [05:14:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18885s) you share file system. [05:14:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18887s) The next emerging thing is agents are collaborating a lot more, and that means that you share [05:14:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18896s) a file and files with your human where you're all collaborating with the same thing. [05:15:03](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18903s) This means file systems that are fast, have consistency, and devx you can reason about. [05:15:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18910s) Something we're working on with modal volumes. [05:15:13](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18913s) The related thing is we're seeing a lot more self-improvement. [05:15:17](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18917s) will save and update memory, they'll run background tasks, which is now called [05:15:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18923s) dreaming, to update this memory over time. [05:15:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18926s) Actually this also just looks like files in a file system. [05:15:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18931s) But again, you want really great performance from this file system. [05:15:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18938s) and the scary thing about agents is they can mess up and they don't really know when they might [05:15:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18948s) leak your sensitive data to a german forum or something so you need to have a lot of control [05:15:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18954s) over what they can and cannot do and you need things like controlling what ips can even talk [05:16:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18961s) to the agent, how do you authenticate it, what domains can it talk to, and also if you're [05:16:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18968s) using credentials for talking to services, basically you want to inject them in a way [05:16:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18974s) that the agent cannot actually see them and then exfiltrate them somewhere later. [05:16:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18980s) And the interesting evolution of this is we've realized that domain-related allow lists are [05:16:27](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18987s) actually not enough. [05:16:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18988s) a lot of our users actually want very specific control [05:16:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18992s) over even like what paths can I talk to [05:16:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18996s) on let's say Datadog or GitHub. [05:16:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=18998s) Like maybe I want to reason about specific operations [05:16:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19002s) that the agent is doing. [05:16:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19003s) And going back to sidecars, [05:16:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19005s) we've seen this emerging as a pattern [05:16:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19007s) where users actually write their own lightweight proxy [05:16:51](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19011s) that has custom rules about what exactly you can do. [05:16:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19013s) And Ramp's another example of a company [05:16:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19017s) that's been using this successfully. [05:17:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19022s) And the final underrated thing is it's really important to have really great observability. [05:17:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19028s) Going back to the example of auto research and an agent doing a lot of training, what [05:17:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19034s) happens when the training run fails? [05:17:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19036s) You want your agent to be able to easily look at the logs of the failure, reason about what [05:17:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19041s) happened and do another iteration with a different set of parameters. [05:17:27](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19047s) While training is running, you want to look at the metrics, GPU utilization and make sure [05:17:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19053s) it's not stalled. [05:17:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19054s) Maybe you want to look at the budget and make sure you're not spending too much on your [05:17:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19058s) sandbox. [05:17:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19060s) And obviously all of this is really important for humans, but as agents are driving a lot [05:17:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19065s) more compute, we're spending a lot of time exposing all of this to the agents themselves [05:17:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19070s) so they can reason about all of this. [05:17:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19074s) All right. [05:17:55](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19075s) I think we know what we need. [05:17:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19077s) Why don't we have Astra build us an agent platform? [05:18:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19081s) Well, I don't think it's a good idea. [05:18:03](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19083s) I don't think you should vibe code a sandbox platform. [05:18:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19086s) And there's three main reasons for that. [05:18:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19088s) Number one is reliability. [05:18:10](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19090s) I think we're all painfully aware how hard it is to get right. [05:18:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19096s) Certain websites we use every day keep going down. [05:18:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19099s) and the reason for that is even with AI, [05:18:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19103s) it's actually really hard to reason about the feedback loop [05:18:26](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19106s) of what happens in the case of this tail event [05:18:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19110s) when the system breaks [05:18:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19111s) and this is why you need to be very careful in building systems. [05:18:35](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19115s) Security is another one. [05:18:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19116s) You could vibe code a sandbox platform [05:18:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19118s) but who is going to do all the work [05:18:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19120s) to continually reason about the security boundaries [05:18:43](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19123s) and make sure it's defensible [05:18:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19127s) in an age when everything is evolving very fast. [05:18:52](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19132s) And finally, scale. [05:18:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19133s) We have a blog post about how we scale to a million sandboxes, [05:18:57](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19137s) which involves a lot of very interesting engineering design [05:19:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19140s) and reasoning about systems. [05:19:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19142s) And this is, again, something that you kind of don't want to leave up [05:19:07](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19147s) to a Vibe-coded platform, [05:19:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19148s) because when you get to scale, you want it to work. [05:19:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19152s) And on that point, we're actually seeing massive scale [05:19:16](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19156s) for a lot of things I mentioned. [05:19:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19158s) So, for example, for the snapshots I mentioned, [05:19:21](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19161s) it's growing quite fast. [05:19:23](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19163s) We're seeing 20 million snapshots a week. [05:19:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19165s) We're processing about eight petabytes of data a day [05:19:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19170s) for each of these snapshots. [05:19:33](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19173s) And our users are creating almost 2.6 million volumes a week, [05:19:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19178s) and all of this is growing super fast. [05:19:42](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19182s) And just to take a customer workload for personal assistant, [05:19:47](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19187s) You need to create 10,000 sandboxes a second. [05:19:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19190s) You need to run almost a million sandboxes. [05:19:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19194s) And they're not going to last very long, they'll just last for like two or three minutes. [05:19:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19198s) But when they come up, [05:20:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19200s) you want them to be interactive in less than a second. That's what this workload really demands. [05:20:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19208s) 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. [05:20:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19219s) So one thing we spend a lot of time doing is actually aggregating cloud capacity from [05:20:24](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19224s) a bunch of hyperscalers and neoclouds. [05:20:27](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19227s) So our users have access to CPU capacity across the world when they need it, but we cannot [05:20:32](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19232s) rely on one cloud provider to do that. [05:20:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19236s) Okay, so where is all this going? [05:20:38](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19238s) I hinted at this earlier, but the harnesses are really good at making use of dynamic environments. [05:20:46](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19246s) and what this is leading to is a lot of separation of the brains, [05:20:51](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19251s) which is the harness, from the hands, [05:20:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19253s) which is the sandbox or the runtime where stuff is running. [05:20:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19258s) And what this means, I have an example here, [05:21:01](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19261s) is the harnesses are going to be able to spin up environments [05:21:05](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19265s) based on what they need right now. [05:21:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19268s) They're good at snapshotting them and keeping them around [05:21:12](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19272s) as artifacts that they need later. [05:21:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19274s) And again, adjusting all of this as necessary. [05:21:18](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19278s) So I think this points at an interesting evolution [05:21:22](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19282s) of the pets to cattle analogy we're all familiar with. [05:21:28](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19288s) To recap, in the olden days, [05:21:31](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19291s) people thought of services like pets. [05:21:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19294s) Each one was bespoke. [05:21:35](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19295s) You had to manage it. [05:21:36](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19296s) You had to reason about this individual deployment specifically. [05:21:41](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19301s) And then over time, with containerization and Kubernetes, people started to think about [05:21:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19308s) it as homogenous templates. [05:21:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19313s) These servers are all the same, you can scale them up and down. [05:21:56](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19316s) If one of them fails, it's not an issue because our system can heal from that. [05:22:00](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19320s) And I think this is where a lot of people deploying agents are today. [05:22:05](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19325s) They're kind of using these homogenous environments and scaling up these agents. [05:22:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19328s) But I think that's really limiting because the models are so good at reasoning about [05:22:14](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19334s) dynamic environments, snapshotting, creating lots of environments when needed and working [05:22:19](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19339s) across them. [05:22:20](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19340s) So what does the new world look like? [05:22:25](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19345s) analogy here is it kind of looks like an ocean of microbes, of all these organisms [05:22:30](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19350s) that are self-evolving and working with each other. [05:22:34](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19354s) credits this fix for the analogy here. [05:22:37](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19357s) but one thing I'm really excited about [05:22:40](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19360s) is as you give agents these really rock-solid primitives [05:22:45](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19365s) for compute and storage and networking, [05:22:48](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19368s) they will be able to actually do a lot more [05:22:50](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19370s) by doing the stuff I talked about, [05:22:53](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19373s) like spinning up lots of environments, [05:22:54](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19374s) working with each other, and being very dynamic. [05:22:58](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19378s) So yeah, I and us at Modal really think [05:23:02](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19382s) that great primitives let your agents do more, [05:23:06](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19386s) and yeah if you're interested in talking about this more [05:23:08](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19388s) reach out to us [05:23:09](https://www.youtube.com/watch?v=akEqiXFo3GM&t=19389s) thank you