00:03:09 · Ruby Justice Thelot, Dexter Horthy
[00:03:09] S1: Yeah. And I know you spent the last time actually talking to a lot of these enterprise customers and like you alluded to, things work great in a demo, but then when you get to production, there are all these issues that enterprise customers are facing that maybe weren't as obvious. Can you tell us a little bit about that? [00:03:26] S0: Well, so multiplayer brings with it a lot of interesting challenges in the enterprise, especially around audit and authentication and access to systems. If I'm talking to my agent, the agent can just act on my behalf and perhaps look at my Slack, edit Notion documents as me, interact with GitHub. But as soon as you have multiple people in a single conversation, you need to kind of compute what is the what is the minimum surface that all those people have access to, or or or assign a specific authentication profile to that session. And then when things are done on behalf of the agent, there's all these audit needs as well. And so it's like how do you how do you actually track and attribute specific actions to a human upstream from what was said and how do you sort of bring that all together. So I I think the the enterprise, and what Factory has done really, really well is build a system with all of these governance and controls and even things like cost tracking, which are not really like a compliance thing but are incredibly useful when you have a large team with hundreds of engineers. So we're excited to kinda bring these ideas together.
00:03:26 · Dexter Horthy, Ruby Justice Thelot
[00:03:26] S0: Well, so multiplayer brings with it a lot of interesting challenges in the enterprise, especially around audit and authentication and access to systems. If I'm talking to my agent, the agent can just act on my behalf and perhaps look at my Slack, edit Notion documents as me, interact with GitHub. But as soon as you have multiple people in a single conversation, you need to kind of compute what is the what is the minimum surface that all those people have access to, or or or assign a specific authentication profile to that session. And then when things are done on behalf of the agent, there's all these audit needs as well. And so it's like how do you how do you actually track and attribute specific actions to a human upstream from what was said and how do you sort of bring that all together. So I I think the the enterprise, and what Factory has done really, really well is build a system with all of these governance and controls and even things like cost tracking, which are not really like a compliance thing but are incredibly useful when you have a large team with hundreds of engineers. So we're excited to kinda bring these ideas together. [00:04:24] S1: Yeah. What's the hardest part about this? You mentioned a lot of things when it comes to security, when it comes to governance, cost tracking. What do you think is the hardest piece from an engineering standpoint?
00:04:24 · Ruby Justice Thelot, Dexter Horthy
[00:04:24] S1: Yeah. What's the hardest part about this? You mentioned a lot of things when it comes to security, when it comes to governance, cost tracking. What do you think is the hardest piece from an engineering standpoint? [00:04:34] S0: Oh, I mean, just the the the the core, you're you're essentially... HumanLayer essentially was a... And and Factory as well is is a giant distributed system with all these different moving that do all do different things and have different different responsibilities. And so building the actual technology layer where something can happen on my workstation and someone else can interact with it, but only in certain capacities. So we we explored a lot of, like, modern technology, like sync engines and durable streams and kind of, like, almost things shaped like a... This idea of, like, conflict conflict free replicated data types where, like, you can have multiple disparate systems making changes to a single item or object, and how do you merge those all together? This is how systems like Google Docs, and Notion and Figma work where you have lots of actions happening in real time, and the system has to take all of this, assume maybe some of it is coming late or coming, twice and being retried, how do you actually maintain one single, you know, source of truth distributed across all these stateful components?
00:04:34 · Dexter Horthy, Ruby Justice Thelot
[00:04:34] S0: Oh, I mean, just the the the the core, you're you're essentially... HumanLayer essentially was a... And and Factory as well is is a giant distributed system with all these different moving that do all do different things and have different different responsibilities. And so building the actual technology layer where something can happen on my workstation and someone else can interact with it, but only in certain capacities. So we we explored a lot of, like, modern technology, like sync engines and durable streams and kind of, like, almost things shaped like a... This idea of, like, conflict conflict free replicated data types where, like, you can have multiple disparate systems making changes to a single item or object, and how do you merge those all together? This is how systems like Google Docs, and Notion and Figma work where you have lots of actions happening in real time, and the system has to take all of this, assume maybe some of it is coming late or coming, twice and being retried, how do you actually maintain one single, you know, source of truth distributed across all these stateful components? [00:05:32] S1: I wanna ask because typically when you're talking about a Notion or you're talking about a Google Doc, yes, you have multiplayers, but it's no more than a dozen at a time. Now when you bring in agent swarms, so this is not just multiplayer agents, these are swarms of agents. How does that change the underlying software and how you have to think about these tools?
00:05:53 · Dexter Horthy
[00:05:53] S0: I mean, we we are always talking about leverage is how do you give humans better leverage on their time. The the slowest way to work with an agent is to ask it to do a thing and it does one thing and then you go review the work all in one without without any upfront planning between the human and the agent. And then as as as you get a little bit more sophisticated in your workflows, you do some upfront planning, maybe these plans get reviewed by multiple people, And the fact that swarms of agents are now able to do, you know, months worth of work in a couple days makes it all the more important that humans can get together and align either one one human with the agent or a team of humans working with that agent. And and the human's job is in software, in teams where there's, you know, regulated industries, we can't just go lights off and ship, you know, a 100,000 lines of slop and hope for the best. Right? If something is wrong, someone's getting fined millions of dollars. There's a need for humans to maintain control and understanding of the system and be the end all decision makers. And now that the scope of work of what agents can accomplish in a given amount of time, the need for alignment and the ability to pull both context and decisions out of the human and also push context from the agent about what was done and why and what needs the most attention during the review phase, is just even more important and even more high leverage.
00:07:09 · Ruby Justice Thelot
[00:07:09] S1: Man, that is so exciting. So today, you officially announced your team is joining Factory and you're open sourcing the platform. What does that mean? What can customers expect to see?