Brains vs Hands: How to Run AI Agents Safely in Production — Viren Baraiya

Class. Thank you to everyone who Class. Thank you to everyone who joined this joined this joined this morning. I know it's morning. I know it's morning. I know it's been a been a been a busy week. Lots of busy week. Lots of busy week. Lots of talk about agents, talk about agents, talk about agents, creating agents, creating agents, creating agents, programming programming programming agents, and all that stuff agents, and all that stuff agents, and all that stuff . And what I . And what I . And what I want to talk about in the want to talk about in the want to talk about in the next 15-20 minutes next 15-20 minutes is launching agents in is launching agents in is launching agents in production. Um, and the things production. Um, and the things that are on everyone's that are on everyone's that are on everyone's mind right now when it comes mind right now when it comes mind right now when it comes to to to building and running building and running building and running agents—that's, you know, agents—that's, you know, agents—that's, you know, tooling. Before we tooling. Before we tooling. Before we start, I start, I start, I would like to ask would like to ask would like to ask the audience, okay? the audience, okay? the audience, okay? How many of you are How many of you are How many of you are launching launching launching agents in production today? agents in production today? Perfectly. So, a lot of Perfectly. So, a lot of hands raised. Um, I hands raised. Um, I hands raised. Um, I was at another was at another was at another conference a few conference a few conference a few weeks ago and weeks ago and weeks ago and asked the same asked the same asked the same question. It seems question. It seems question. It seems there was only one hand, there was only one hand, there was only one hand, so that's pretty good. so that's pretty good. so that's pretty good. It seems like, you know, everything is It seems like, you know, everything is It seems like, you know, everything is moving forward. Em. moving forward. Em. OK. So when we OK. So when we think about agents, think about agents, think about agents, right? If we right? If we right? If we go back to the go back to the go back to the equivalents of "Hello World" equivalents of "Hello World" equivalents of "Hello World" for agents, uh, it for agents, uh, it for agents, uh, it always starts with, always starts with, always starts with, you know, an agent you know, an agent you know, an agent calling a tool calling a tool , and how well the LLM , and how well the LLM , and how well the LLM decides to call the decides to call the decides to call the tool, maybe tool, maybe tool, maybe using some using some using some context or memory context or memory context or memory and planning everything in advance, and planning everything in advance, and planning everything in advance, right? Um, what we're right? Um, what we're right? Um, what we're starting starting starting to realize is that, you to realize is that, you to realize is that, you know, it's a very know, it's a very know, it's a very small, narrow small, narrow small, narrow picture, if picture, if picture, if you think about it. When you you think about it. When you you think about it. When you think of agents in think of agents in think of agents in production, it's not production, it's not production, it's not always just chatbots.

always just chatbots. always just chatbots. Um, the early examples Um, the early examples Um, the early examples we saw were like, we saw were like, we saw were like, we're we're we're going to build a going to build a customer service chatbot or something customer service chatbot or something like that. Um, but when like that. Um, but when like that. Um, but when we start thinking we start thinking we start thinking about agents more broadly, about agents more broadly, about agents more broadly, think about think about think about background agents, uh, background agents, uh, background agents, uh, background workers, background workers, background workers, agents who work agents who work agents who work on a schedule. Um, I'd on a schedule. Um, I'd on a schedule. Um, I'd like to have an agent like to have an agent like to have an agent that runs that runs that runs every hour to every hour to every hour to check, uh, what's check, uh, what's check, uh, what's going on with my going on with my going on with my schedule and see if there's schedule and see if there's schedule and see if there's anything new anything new anything new that I should that I should that I should know about. Uh, event- know about. Uh, event- driven agents, driven agents, driven agents, right? Agents that right? Agents that right? Agents that listen for various listen for various listen for various events. Um, to give you an events. Um, to give you an events. Um, to give you an example, I have a example, I have a example, I have a work system where work system where work system where my notifications are my notifications are my notifications are triggered every now and triggered every now and triggered every now and then. I have logs then. I have logs then. I have logs coming through OTEL. I would coming through OTEL. I would coming through OTEL. I would like my like my like my agents to listen to these agents to listen to these agents to listen to these events, react to events, react to events, react to them, and see what's them, and see what's them, and see what's going on there, right? Um, going on there, right? Um, going on there, right? Um, I can have I can have I can have long-term long-term long-term coordinators, and by coordinators, and by coordinators, and by that I mean I have an that I mean I have an that I mean I have an agent who agent who agent who works for a works for a works for a long time. I would long time. I would long time. I would like an agent to like an agent to like an agent to monitor another monitor another monitor another agent to see, " agent to see, " Hey, what's Hey, what's Hey, what's going on there and going on there and going on there and should I should I should I push this push this push this agent to move on agent to move on agent to move on or take other or take other or take other actions?" And everything is in actions?" And everything is in actions?" And everything is in that spirit, right? Um, one that spirit, right? Um, one that spirit, right? Um, one more point more point more point is about is about is about multi-agent multi-agent multi-agent systems, right? Agents systems, right? Agents systems, right? Agents communicating with communicating with communicating with other agents.

other agents. other agents. The last thing you The last thing you The last thing you want is for one want is for one want is for one agent to be responsible for agent to be responsible for agent to be responsible for many things, many things, many things, risking risking risking hallucinations. Ahem. hallucinations. Ahem. So the more we So the more we think about agents, think about agents, think about agents, right? I'd like to right? I'd like to right? I'd like to think of an agent as think of an agent as think of an agent as something that becomes more like an something that becomes more like an something that becomes more like an application, application, application, rather than just a component rather than just a component , right? Um, if , right? Um, if , right? Um, if we go back to the we go back to the we go back to the old school old school old school of microservices, when of microservices, when of microservices, when we were building and probably we were building and probably we were building and probably still are building still are building still are building microservices, right? You microservices, right? You microservices, right? You wouldn't release wouldn't release a microservice that a microservice that does everything into production, right? At this does everything into production, right? At this does everything into production, right? At this point, it ceases point, it ceases point, it ceases to be a microservice. to be a microservice. to be a microservice. We usually We usually We usually release a lot of release a lot of release a lot of microservices. They microservices. They microservices. They communicate with each communicate with each communicate with each other through other through other through choreography or choreography or choreography or orchestration, and you orchestration, and you orchestration, and you have an application have an application have an application that relies on a that relies on a that relies on a set of microservices set of microservices set of microservices that essentially that essentially that essentially follow a follow a follow a single single single responsibility principle to responsibility principle to responsibility principle to achieve business achieve business goals. There's a very goals. There's a very goals. There's a very clear parallel to this clear parallel to this when it comes to when it comes to when it comes to agent shells, right agent shells, right ? If you think about ? If you think about ? If you think about a shell that is a shell that is a shell that is responsible for responsible for responsible for achieving a achieving a achieving a specific goal.

specific goal. specific goal. Let's say I have a Let's say I have a Let's say I have a shell to shell to shell to do my do my do my SRE work. Um, she's SRE work. Um, she's SRE work. Um, she's working with a few working with a few working with a few agents to see agents to see agents to see what's going on there, what's going on there, what's going on there, right? Maybe it's right? Maybe it's right? Maybe it's monitoring events from monitoring events from monitoring events from my logging system, my logging system, my logging system, communicating with my communicating with my communicating with my Kubernetes cluster as an Kubernetes cluster as an Kubernetes cluster as an agent to find out agent to find out agent to find out what's going on there. what's going on there. Um, my metric Um, my metric systems, my systems, my customer service dashboards, customer service dashboards, and so on. Um, and I and so on. Um, and I and so on. Um, and I want to emphasize that want to emphasize that want to emphasize that if if if you think about it, right? you think about it, right? you think about it, right? Your shell is the Your shell is the Your shell is the application, and vice versa application, and vice versa , right? One agent doesn't , right? One agent doesn't , right? One agent doesn't provide the provide the provide the entire design, but entire design, but entire design, but when you start when you start when you start putting everything together and putting everything together and putting everything together and building a wrapper that building a wrapper that building a wrapper that controls the execution of the controls the execution of the controls the execution of the agent, implementing the agent, implementing the agent, implementing the design, that's when you design, that's when you design, that's when you start thinking about start thinking about start thinking about an application, right? These are an application, right? These are an application, right? These are systems like your systems like your systems like your databases, internal databases, internal databases, internal systems, systems, systems, corporate systems corporate systems . These are people, a person in the . These are people, a person in the . These are people, a person in the control loop. Um control loop. Um , these are tools , these are tools , these are tools via API or MCP via API or MCP via API or MCP working together. Um, and working together. Um, and working together. Um, and as I mentioned, right as I mentioned, right ? The shell runs ? The shell runs ? The shell runs more than just an more than just an more than just an agent loop. And this is an agent loop. And this is an agent loop. And this is an important thing I important thing I important thing I realized while realized while realized while trying trying trying to create wrappers to create wrappers to create wrappers for real for real for real production scenarios production scenarios . It's a combination of . It's a combination of . It's a combination of deterministic and deterministic and deterministic and non-deterministic non-deterministic non-deterministic parts of the application, parts of the application, parts of the application, right?

right? Nondeterminism Nondeterminism is provided by LLM in is provided by LLM in is provided by LLM in terms of reasoning, in terms of the terms of reasoning, in terms of the thought process. Um, and then there's the thought process. Um, and then there's the nondeterministic nondeterministic nondeterministic part, and there's the part, and there's the part, and there's the deterministic deterministic deterministic part, right? Where do you part, right? Where do you part, right? Where do you not want not want non-determinism to leak, non-determinism to leak, right? Think right? Think right? Think about payments. Um, or about payments. Um, or about payments. Um, or let's say if I have a let's say if I have a let's say if I have a harness agent that's harness agent that's harness agent that's responsible for responsible for responsible for monitoring my monitoring my monitoring my production deployment production deployment , how I manage my , how I manage my , how I manage my Kubernetes clusters, I Kubernetes clusters, I Kubernetes clusters, I probably want to have a probably want to have a probably want to have a very well- very well- very well- defined defined defined workflow for what workflow for what workflow for what sequence of steps sequence of steps I follow when I I follow when I I follow when I want to restart want to restart want to restart my cluster. And it's a my cluster. And it's a my cluster. And it's a very deterministic very deterministic very deterministic set of processes that set of processes that set of processes that I know exactly what it's doing every time I know exactly what it's doing every time I know exactly what it's doing every time . . . So I want my So I want my So I want my harness to provide harness to provide harness to provide determinism when it determinism when it determinism when it matters. Um, and of course, matters. Um, and of course, matters. Um, and of course, harness—it's a harness—it's a harness—it's a long process long process , right? It works , right? It works , right? It works for a certain for a certain for a certain time. It can time. It can time. It can run from run from run from a few seconds if a few seconds if a few seconds if it's something really quick, it's something really quick, it's something really quick, like "hey, like "hey, like "hey, check what's check what's check what's going on here," to going on here," to going on here," to days, months, or days, months, or days, months, or even longer, right even longer, right even longer, right ? Think about ? Think about ? Think about long-running processes long-running processes , order management systems , order management systems , order management systems where you where you where you wait for third-party wait for third-party wait for third-party services to services to services to deliver your deliver your deliver your shipment, or shipment, or shipment, or wait for people wait for people wait for people to take action and to take action and to take action and agree on things, and so agree on things, and so agree on things, and so on. Or is Harness just on. Or is Harness just on. Or is Harness just waiting, right? He waiting, right? He waiting, right? He waits for waits for waits for events to happen.

events to happen. events to happen. So, when this event So, when this event So, when this event happens, you take happens, you take happens, you take action and do something action and do something action and do something around it. So, that around it. So, that around it. So, that brings up another brings up another brings up another important point, important point, important point, right? That when you right? That when you right? That when you think about think about think about long-term systems long-term systems , you need , you need , you need reliability, right? reliability, right? reliability, right? You want to be You want to be You want to be able to able to able to recover when recover when recover when something goes wrong something goes wrong something goes wrong because ultimately these because ultimately these because ultimately these harnesses are running somewhere harnesses are running somewhere harnesses are running somewhere throughout your throughout your throughout your infrastructure infrastructure infrastructure stack, right? stack, right? stack, right? Maybe in the cloud, Maybe in the cloud, Maybe in the cloud, maybe in a sandbox maybe in a sandbox , but these things can , but these things can , but these things can rise and rise and rise and fall. There could be fall. There could be fall. There could be network failures, network failures, network failures, separations, anything separations, anything separations, anything that could happen. So, that could happen. So, that could happen. So, you need the harness to you need the harness to you need the harness to be reliable. And be reliable. And be reliable. And one thing about one thing about one thing about reliability here—it's a reliability here—it's a reliability here—it's a must, must, must, right? Ultimately, right? Ultimately, right? Ultimately, reliability is the price of reliability is the price of reliability is the price of entry. This is not the entry. This is not the entry. This is not the feature you are feature you are feature you are looking for in harness. Um, and as looking for in harness. Um, and as looking for in harness. Um, and as I mentioned, I mentioned, I mentioned, right? The loop in which right? The loop in which right? The loop in which harness operates encompasses harness operates encompasses harness operates encompasses the agent and everything else. So the agent and everything else. So the agent and everything else. So now let's now let's now let's think about how think about how think about how these harnesses work, shall we these harnesses work, shall we these harnesses work, shall we ? Um, if you're ? Um, if you're ? Um, if you're thinking about a harness, right?

thinking about a harness, right? thinking about a harness, right? And the cycle, essentially what it And the cycle, essentially what it And the cycle, essentially what it does—it has a state of the does—it has a state of the does—it has a state of the world. He knows what world. He knows what world. He knows what work has been done. work has been done. work has been done. Um, what was written Um, what was written Um, what was written about the side about the side about the side effects, correct? effects, correct? effects, correct? So, for example, if So, for example, if So, for example, if I did a I did a I did a cluster restart, I know that cluster restart, I know that cluster restart, I know that it's fixed. Um, it's fixed. Um, it's fixed. Um, if I sent an if I sent an if I sent an email, I email, I email, I know it happened. know it happened. know it happened. Um, I know what Um, I know what Um, I know what worked and what didn't. And worked and what didn't. And worked and what didn't. And then, based on the then, based on the then, based on the current state of current state of current state of the world and the goal, I know the world and the goal, I know the world and the goal, I know what needs to happen next what needs to happen next , right? And this is where , right? And this is where reasoning and LLM come into play. Um, and what's reasoning and LLM come into play. Um, and what's actually happening here is that actually happening here is that if you think about the if you think about the clear distinction, which clear distinction, which clear distinction, which is the most important thing. If there's is the most important thing. If there's one thing I want you all to take away from this slide, it's this: the responsibility of a the responsibility of a nondeterministic nondeterministic nondeterministic agent is to agent is to agent is to plan what plan what plan what happens next, not happens next, not happens next, not to take action. Um, and to take action. Um, and to take action. Um, and then Harness is the one who then Harness is the one who then Harness is the one who directly directly directly does the work. And that's does the work. And that's does the work. And that's important for a variety of important for a variety of important for a variety of reasons, right? reasons, right?

One of them is that One of them is that Harness is a Harness is a Harness is a deterministic deterministic deterministic piece of code that piece of code that piece of code that really understands what it really understands what it really understands what it means to do a means to do a means to do a certain job. So, certain job. So, certain job. So, as I mentioned, as I mentioned, as I mentioned, if I'm trying to if I'm trying to if I'm trying to create a Harness that create a Harness that create a Harness that runs my DevOps or runs my DevOps or runs my DevOps or SRE agents, when the agent SRE agents, when the agent SRE agents, when the agent says, "Hey, this says, "Hey, this says, "Hey, this cluster is down, cluster is down, cluster is down, it should be it should be it should be restarted." restarted." Harness has to decide how to Harness has to decide how to reboot reboot reboot the cluster, what it the cluster, what it the cluster, what it takes to do that, and that has to takes to do that, and that has to takes to do that, and that has to be, at least in be, at least in be, at least in my understanding, a very my understanding, a very my understanding, a very deterministic set of deterministic set of deterministic set of processes, right? processes, right? processes, right? So that every So that every So that every cluster restart is cluster restart is cluster restart is exactly the same. There's exactly the same. There's exactly the same. There's no no no other category here, other category here, other category here, right? Um, also, right? Um, also, right? Um, also, if if if approval is needed, like approval is needed, like approval is needed, like if I'm trying to if I'm trying to if I'm trying to reboot a reboot a reboot a production cluster production cluster , I probably want to , I probably want to , I probably want to have human have human have human control, maybe control, maybe control, maybe send send send a message in Slack to a message in Slack to a message in Slack to someone saying, " someone saying, " Hey, I'm going to Hey, I'm going to Hey, I'm going to reboot this reboot this reboot this cluster. Do you cluster. Do you cluster. Do you think that's okay think that's okay think that's okay or not, right?" And I don't or not, right?" And I don't or not, right?" And I don't want it to depend want it to depend want it to depend on the hallucinations of LLMs on the hallucinations of LLMs on the hallucinations of LLMs who might say who might say who might say it doesn't it doesn't it doesn't need to be done. This must be need to be done. This must be need to be done. This must be guaranteed at the guaranteed at the guaranteed at the execution level. So everything execution level. So everything execution level. So everything has to be very has to be very has to be very deterministic when it comes to deterministic when it comes to things like this. Um, ideally I things like this. Um, ideally I want it to be want it to be want it to be idempotent, and if idempotent, and if idempotent, and if not, I want it to be not, I want it to be not, I want it to be recorded that it's not recorded that it's not recorded that it's not idempotent and what idempotent and what idempotent and what exactly happened. So I exactly happened. So I exactly happened. So I can take care of the can take care of the can take care of the side effects or side effects or side effects or treat them separately, treat them separately, treat them separately, right? Um, so that's right? Um, so that's right? Um, so that's the most important thing, the most important thing, the most important thing, right? A clear right? A clear right? A clear division into what we, as it were division into what we, as it were division into what we, as it were , call, right?

, call, right? , call, right? Brain and hands. Harness is the Brain and hands. Harness is the Brain and hands. Harness is the hands, and the brain is the LLM. hands, and the brain is the LLM. hands, and the brain is the LLM. Ahem. So, yes, if the Ahem. So, yes, if the Ahem. So, yes, if the agent writes a plan, then agent writes a plan, then agent writes a plan, then Harness executes that plan. Harness executes that plan. And I'll show you a And I'll show you a quick quick quick demonstration of how it demonstration of how it demonstration of how it all comes all comes all comes together. Hmm, but is this a together. Hmm, but is this a together. Hmm, but is this a new concept? new concept? new concept? If you think about it, this isn't If you think about it, this isn't If you think about it, this isn't necessarily a new necessarily a new necessarily a new concept. This has been around for concept. This has been around for concept. This has been around for a while, right? a while, right? a while, right? If we consider If we consider If we consider workflows as " workflows as " sagas" that we used to sagas" that we used to sagas" that we used to write ourselves and clearly write ourselves and clearly write ourselves and clearly define what is define what is define what is happening. It was happening. It was happening. It was and is a very and is a very and is a very deterministic set of deterministic set of deterministic set of processes. When you processes. When you processes. When you think about the think about the think about the Harness agents, they Harness agents, they Harness agents, they are essentially " are essentially " late-link" late-link" late-link" sagas. And what I sagas. And what I sagas. And what I mean is that they mean is that they mean is that they have a very limited have a very limited have a very limited and deterministic and deterministic and deterministic set of tools set of tools set of tools and tasks that they and tasks that they and tasks that they can work with. can work with. Instead of Instead of assembling them assembling them assembling them in advance, the agent in advance, the agent in advance, the agent proposes and builds them proposes and builds them proposes and builds them at runtime. And at runtime. And at runtime. And so, essentially, they are so, essentially, they are so, essentially, they are late-connected late-connected late-connected sagas, right? But you sagas, right? But you sagas, right? But you get all get all get all the benefits of a traditional the benefits of a traditional the benefits of a traditional saga saga saga workflow out of the box: workflow out of the box: workflow out of the box: visibility into what's visibility into what's visibility into what's happening, happening, happening, the ability the ability the ability to control everything and to control everything and to control everything and iterate on steps, iterate on steps, iterate on steps, such as how such as how such as how far into the future far into the future far into the future an agent can plan.

an agent can plan. You can plan You can plan one step at one step at one step at a time or multiple a time or multiple a time or multiple steps at once. And so, if steps at once. And so, if steps at once. And so, if you think about it, right? It's you think about it, right? It's you think about it, right? It's more like a more like a more like a branched branched branched workflow. If only it workflow. If only it workflow. If only it were possible were possible were possible to build a to build a to build a workflow with every workflow with every workflow with every possible combination of possible combination of possible combination of branches, then, you know, branches, then, you know, branches, then, you know, it would build it itself, it would build it itself, it would build it itself, unlike unlike unlike agents that agents that agents that simplify the work. simplify the work. If you have n- number of number of number of tools, the agent tools, the agent tools, the agent can perform n can perform n can perform n different combinations, which would different combinations, which would different combinations, which would otherwise be almost otherwise be almost otherwise be almost impossible impossible impossible to predict to predict to predict in advance and in advance and in advance and implement. So implement. So implement. So let's take a look, let's take a look, let's take a look, okay? I'm going to okay? I'm going to okay? I'm going to quickly show the quickly show the quickly show the completed launch. completed launch. What I was talking about What I was talking about earlier, yes. Here's one earlier, yes. Here's one earlier, yes. Here's one of my examples of an of my examples of an of my examples of an agent that essentially agent that essentially agent that essentially acts as an SRE agent, yes. acts as an SRE agent, yes. His task is to understand what is to understand what is to understand what is happening in the happening in the happening in the current system, current system, current system, try try try to plan the next to plan the next to plan the next steps, and execute them.

steps, and execute them. So it's essentially a So it's essentially a fix loop that fix loop that fix loop that runs twice. runs twice. runs twice. During the first During the first During the first iteration, it iteration, it iteration, it tries tries tries to understand to understand to understand the root cause of the root cause of the root cause of the problem, the problem, the problem, tries to...react tries to...react tries to...react to it, observes to it, observes to it, observes its output, and then its output, and then its output, and then runs another loop and runs another loop and runs another loop and plans a new set of plans a new set of plans a new set of steps to see steps to see steps to see what needs to happen next what needs to happen next , right? So, as , right? So, as , right? So, as input for my agent, input for my agent, input for my agent, let's see let's see let's see what exactly was entered what exactly was entered what exactly was entered here. Step number one here. Step number one is, you know, the call for is, you know, the call for is, you know, the call for LLM. As you can see, this is an SRE LLM. As you can see, this is an SRE agent. Of course, this is a demo agent. Of course, this is a demo , so everything , so everything , so everything is planned is planned is planned in advance to some extent. The LLM output, in advance to some extent. The LLM output, in advance to some extent. The LLM output, as you can see, is as you can see, is as you can see, is the steps it needs the steps it needs the steps it needs to take. What he's to take. What he's to take. What he's saying is that, saying is that, saying is that, as you can see as you can see , instead of , instead of , instead of just doing one step at a just doing one step at a just doing one step at a time, he's essentially time, he's essentially time, he's essentially suggesting a suggesting a suggesting a sequence of actions. sequence of actions.

You should collect You should collect evidence, evidence, evidence, analyze logs, analyze logs, analyze logs, and, if necessary, and, if necessary, roll back the roll back the deployment and deployment and deployment and verify verify verify recovery based on the collected evidence. This is then recovery based on the collected evidence. This is then recovery based on the collected evidence. This is then passed to a passed to a passed to a specialized specialized specialized tool tool tool called "planning and called "planning and called "planning and compilation". So, this is the compilation". So, this is the compilation". So, this is the plan that the plan that the plan that the agent provided. It agent provided. It agent provided. It compiles into a very compiles into a very compiles into a very deterministic deterministic deterministic workflow. We workflow. We workflow. We rely on rely on rely on Conductor here as the Conductor here as the Conductor here as the workflow execution engine workflow execution engine workflow execution engine , so , so , so the output is a fully the output is a fully the output is a fully ready-to-run ready-to-run ready-to-run workflow. It's workflow. It's workflow. It's being done, and I'll being done, and I'll show you very quickly what it show you very quickly what it looks like now. Ahem. So, that's the looks like now. Ahem. So, that's the looks like now. Ahem. So, that's the first step of first step of first step of execution, right? If execution, right? If execution, right? If we look here, we look here, we look here, like I said, right? like I said, right? We analyzed We analyzed the logs, reviewed the the logs, reviewed the the logs, reviewed the metric queries, metric queries, metric queries, decided to decided to decided to rollback and test rollback and test rollback and test the recovery, and everything the recovery, and everything the recovery, and everything else. So, step number else. So, step number else. So, step number one is complete. It one is complete. It one is complete. It starts another cycle. starts another cycle. Hmm, the same thing is Hmm, the same thing is happening here.

happening here. happening here. Ahem. In the next Ahem. In the next Ahem. In the next iteration, he decides: iteration, he decides: iteration, he decides: okay, it looks like these okay, it looks like these okay, it looks like these things are already done. things are already done. things are already done. Let's check the Let's check the Let's check the further processes and, further processes and, further processes and, if necessary, if necessary, if necessary, confirm confirm confirm the recovery and the recovery and the recovery and complete it, okay? Um, complete it, okay? Um, complete it, okay? Um, and then the and then the and then the next loop starts and next loop starts and next loop starts and completes the task. completes the task. completes the task. But here's an example of a cycle But here's an example of a cycle that's self-contained, that's self-contained, that's self-contained, sort of " sort of " self-scheduling," right? And self-scheduling," right? And self-scheduling," right? And to clearly show to clearly show to clearly show you what's you what's you what's going on here—it's a going on here—it's a going on here—it's a cycle. Ahem. At each cycle. Ahem. At each cycle. Ahem. At each step along the way, it step along the way, it step along the way, it essentially looks at the essentially looks at the essentially looks at the current state of the current state of the current state of the system. So, this is my system. So, this is my system. So, this is my agent loop that agent loop that agent loop that understands the world, plans understands the world, plans understands the world, plans and executes not one and executes not one and executes not one step at a time, but plans step at a time, but plans step at a time, but plans several steps at once, several steps at once, several steps at once, executes them, checks executes them, checks executes them, checks and completes the loop. Ahem. and completes the loop. Ahem. and completes the loop. Ahem. Now you can Now you can Now you can run this in a run this in a run this in a production production production environment as many environment as many environment as many times as you want. This times as you want. This times as you want. This can run on can run on can run on events or on a events or on a events or on a schedule. Um, like I schedule. Um, like I schedule. Um, like I said, right? So said, right? So said, right? So all of this can all of this can all of this can work for much work for much work for much longer than just a longer than just a longer than just a short period of short period of short period of time. OK. So, it's getting time. OK. So, it's getting time. OK. So, it's getting closer to closer to closer to completion. This is kind of a completion. This is kind of a completion. This is kind of a review of what we review of what we review of what we just did. As just did. As just did. As you can see, during the you can see, during the you can see, during the second iteration we second iteration we second iteration we decided not to do a decided not to do a decided not to do a rollback since it rollback since it rollback since it was already done in the was already done in the was already done in the first iteration. But first iteration. But first iteration. But yes, it's about yes, it's about yes, it's about bindings and agents bindings and agents bindings and agents and how they and how they and how they provide provide provide determinism to

determinism to determinism to your productive your productive your productive application. The example application. The example application. The example I showed you I showed you I showed you works on Conductor. Conductor works on Conductor. Conductor is an is an open open source workflow orchestration platform that supports the source workflow orchestration platform that supports the source workflow orchestration platform that supports the creation of agent creation of agent creation of agent loops as well as loops as well as loops as well as agent systems. It is agent systems. It is agent systems. It is completely open completely open completely open source. We at Orkes source. We at Orkes source. We at Orkes provide an enterprise provide an enterprise provide an enterprise version, but you can version, but you can version, but you can try it out try it out try it out yourself. Here is the QR code. yourself. Here is the QR code. yourself. Here is the QR code. Try it. Try it. Try it. Join Join Join our Slack and if our Slack and if our Slack and if you have any you have any you have any questions, we'll be questions, we'll be questions, we'll be happy to help. happy to help. happy to help. We have a booth here at We have a booth here at We have a booth here at Orkes. Come in if Orkes. Come in if Orkes. Come in if you want to see a you want to see a you want to see a live demo. Whether you live demo. Whether you live demo. Whether you want to run want to run want to run your own LangChain, your own LangChain, your own LangChain, OpenAI, or any OpenAI, or any OpenAI, or any other agents, Conductor can other agents, Conductor can other agents, Conductor can handle it all. handle it all. handle it all. Thank you.