Superflow: AI Workflows With Real Reliability

Added: 2026-05-28

[00:00:00]Hey everyone, my name is Srihari Kulkarni and I lead uh inferencing and orchestration at Lizer. Today I'm going to walk you guys through Superflow. It's Lizer's orchestration platform that is durable and replayable. Uh the two things that are currently missing in the existing um workflow systems. So, what do I mean by durable? Well, even before that, let's go to the first principles.

[00:00:19]Why do you even need a workflow system in the first place? So, um there are two parts to this. One is it should be a low code or no code tool so that non-developers and devs alike can focus on building the business logic part of things, while not worrying about the infrastructure that runs that business logic. So, this is the first principles on why the workflow systems were even built in the first place. So, Superflows handle this both of these parts. So, the infra part of it, it's completely durable and replayable. Uh Superflow is powered by a re-state.dev, which is um durable execution layer that sits below Superflow. Let's get into Superflow. I'm going to create a new Superflow here, and let's say this is the demo Superflow. Uh there are a bunch of templates you can choose from, but for now let's just create an empty one. The use case uh we are going to be building today is just sending out an email with a human in the loop. So, the AI node will generate an email and it will send out uh send it out back to me. But, uh before it sends it out, uh we'll get an uh approval node and you need to we need to check the what the AI has generated and then approve it.

[00:01:26]So, let's drop the trigger node. This is a mandatory node to start any flow.

[00:01:32]There are a bunch of options to trigger uh Superflow. You can enable webhook triggers, you can schedule a trigger, or you can upload a file to run the Superflow. Schedule triggers are pretty self-explanatory. You can provide a frequency, or if you are well-versed with cron uh expressions, you can directly give the cron expressions as well. So, let's say um you wanted to run something every Monday at 9:00 a.m. uh in America New York time. You can choose that over here and just enable the schedule trigger.

[00:02:02]Similarly, you have webhook triggers wherein if you enable it and generate a secret and just call this girl through your external service, it will start the super flow. But for today's use case, let's just run it manually over here in the super flow UI.

[00:02:15]So, this is the trigger node. Let's follow it up with an LLM node. There are default prompts given in, so you don't you don't even have to configure anything. Let's say I'm not even touching anything else. I'm just directly running this.

[00:02:27]I'm saying draft an email explaining the history of video games. Just this much. That's my input. You can see that it calls the LLM node and we'll get an output. So, there you go. It gave me an output with the subject and with the email body as well. So, but this is not good enough, right? It's just a single text continuous text. So, what you can do is you can add output fields over here and you can say the subject. This is structured output from the LLM. You can have a subject and you can have an email body.

[00:03:01]Both of which are strings.

[00:03:03]And then if you are already in the Laser ecosystem, you know that we support a plethora of tools, some of which are connected to my account here and I'm going to choose the Gmail tool connected by a Compozio and I want to send out an email. Here you can see all the values that are needed are pre-populated. So, the only thing I will mention here is let's say the recipient email, which is me, and the body that the LLM generates. Now, you might have a question, how do I extract the body generated from the LLM node into the tool node?

[00:03:36]So, we make it pretty easy here.

[00:03:38]In the body string, you just have to insert the reference from the LLM node.

[00:03:43]You can see that we had provided what structure the output must be in. So, this is the email body and also the subject as well. So, this is supposed to be a string with output.subject. And that's it. You can just save this. Let's run it. Uh I know we said we'll be implementing human in the loop. I'll come to that as well. For now, just run this.

[00:04:06]And there we go. Uh the email was sent.

[00:04:09]So, uh here's the email that was sent from Superflow. Let's uh look at the human in the loop at this point. So, we don't want to directly send out the email. We want to gate it for a human approval and then send it. So, let's do that. Let's drop a wait for approval node and auto align it.

[00:04:27]So, wait for approval and uh we can have an approval message here or uh just leave it empty for the default approval message.

[00:04:35]And uh that that's it actually. You can just run it again.

[00:04:40]And this time, before sending out the email, it will wait for your approval as soon as it reaches this node.

[00:04:47]So, there you go. Now, you might be thinking when it is waiting for approval, there's a lot of flaky systems out there that will just break down while waiting.

[00:04:55]So, Superflow guarantees that you will have zero compute cost when it's waiting for your approval. It It is durable.

[00:05:01]It's just a file on disk, so it will have zero compute cost. It will never forget that there was a waiting approval uh node. So, you can just pause the execution at this point, come back a year or two later, and it'll still resume from the same point. So, uh for now, I can directly approve this.

[00:05:16]There's a bunch of um options on the human in the loop node itself. We'll have a separate video for that. Yeah, uh that's it. I get the email again. So, there we go.

[00:05:25]Okay, um the last part of this video is what if you don't want to configure the tool node to, you know, just insert the reference from the LLM node into the tool node so that it is sent out. Well, because this is something the LLM can do by itself. So, we have an option for that as well. This is in programming what we call the react loop or the agentic loop where the LLM decides what tool needs to be called and it does call it directly. So, let's do that. So, in the tool node, you can see that there is a run as tool call option on the top here.

[00:05:56]Um that [music] same option will uh be available even for other LLM nodes and AI agent nodes. So, if you click on any other LLM or AI agent node and connect it to an LLM upstream like I do here.

[00:06:11]So, you can see that there is a run as sub agent option over here. So, the LLM, the master LLM, the manager agent in this case, decides when to call the sub agent or a sub tool and then execute it by itself without ever requiring you to wire them up. Uh but, of course, you can pause the execution with a work wait for approval node as well. So, for now, um just to demo how this works, I'll run this uh Gmail tool call uh as a sub agent for the manager, that is the first LLM node over here.

[00:06:41]Once we toggle this on, you can see that the edge turns purple. This is the indication that this is now hooked up as a tool in the first LLM's agent tech loop. So, now the agent has authority to delegate the task to the tool node uh whenever it wants to. Now, you might be a bit curious on how do we handle cases where um you want a deterministic input to be given to the LLM node because LLM might hallucinate sometimes. Great question.

[00:07:07]So, in that case, go to the fixed inputs and you'll see all the available uh tool parameters out of which, let's say the recipient email is deterministic. You already know who needs to get the email.

[00:07:19]So, in this case, it's me.

[00:07:20]So, I I just input that. So, uh what does fixed inputs do? The LLM just ignores it completely. It doesn't see uh recipient email anymore while it is delegating the tool call. So, fixed inputs must be used whenever there is a deterministic element involved in your super flow. So, that whenever the LLM decides to delegate a task to the sub agent or a sub tool, It knows that it it need not um generate content for the fixed inputs. It will only generate the content for these arguments down here.

[00:07:50]So, you don't even have to use this anymore. Let me reset that.

[00:07:55]Let's send email.

[00:07:58]Okay, now you see that all the parameters are back to their defaults.

[00:08:02]Save [music] that. Let's run it again.

[00:08:05]Okay, it did not call the tool because I never asked it to send out an email.

[00:08:11]That's why we have to not just say draft an email explaining. We'll also say and send out an email.

[00:08:21]>> [snorts] >> Now it knows that it has to delegate the tool call.

[00:08:26]So, there you go. The LLM decided that it it needs to delegate this task to this tool node over here. And because it is gated with a human in the loop node, it is asking for my approval. Let me just go ahead and approve this.

[00:08:37]And now it calls the tool call.

[00:08:40]The tool node to send out an email. I will be receiving another email.

[00:08:44]There you go.

[00:08:45]And yeah, the LLM just outputs what we had given it as a structured output before, that is email body and subject. And I think that's it for this video. In the next video, we'll be covering how the human in the loop node works, what are its configurations, and what all you can do with it. Thank you.