Introduction

I will talk today about AI agents and workflow automation. And I will mostly show you some examples, so I don't have much slides. But I want to say that first, ask me any questions.

It's very easy. Fuzzy topic, like nobody understands what are AI agents.

But basically, if you're a startup and doing workflows and you rename yourself to agents, your valuation goes up. So it's very complicated. But at the same time, it's quite straightforward.

Speaker and Context

Yeah, my name is Nikolai. I'm founder and CEO of Hyperscale.

Me and Nate, who will speak last today, are also members of Sunday Hacker Club at MIT. So we meet every Sunday and build AI apps.

And I also teach here and there at MIT Harvard courses.

What Are AI Agents?

First of all, I want to explain you what is AI agents. There is no strict definition, and there are multiple opinions.

But basically, if you have LLM, like GPT, for example, or GROK, or CLOD, and you have a prompt, and it produces you some output, that's a usual usage of AI.

From LLMs to Augmented LLMs

But we can define augmented LLM block as the same but which can use tools and have memory.

Usually, AI doesn't have memory. You ask something, it give you an answer, but it don't remember your previous different chats.

Some products like ChatGPT are starting implemented memory. So actually, ChatGPT is not AI model, but AI product with this components also.

Memory, Tools, and Retrieval

More complex using of LLMs, you can use memory, you can use tools, and retrieval, which is RAC. You can access some databases or, for example, when you do deep research, LLM can go online, search for additional information, retrieve it, and put it into your chat, basically.

Autonomy and Feedback Loops

and especially autonomous agents, are basically these augmented LLM blocks which interact with environment and get some feedback.

So you as a human ask agent to do something. Agent go to and do some actions, get a feedback from environment.

It can loop. Sometimes agents are different than workflows by having loops. So agents can work more autonomously and think and reason more autonomously.

uh until the task is done and the smarter and smarter ai is becoming basically agents are becoming more and more autonomous and can do more complicated tasks by themselves without requiring human in the loop

Is it kind of clear? Okay, nice.

Tools for Workflow Automation

So I want you to show two tools today.

One is NI10. There is similar one, Make. Also Stack AI is kind of similar. Zapier is kind of moving into this direction.

So it's a set of tools where you can automate stuff and build the workflows.

n8n Overview and Company Use Cases

It's open source, so you can install it locally, or you can use it as a service. And I will show you some workflows which we actually have in my company. So we are Um, I hope it will not have any sensitive data, but, um, so there's our workflows.

We have admin team. So we are like 40, 50 people, uh, start up and we have admin team and we, and they are automating themselves with, uh, an item.

Uh, yeah. Yeah. Yeah.

So this is our private installation of an item, which is open source. And you can do it, or you can buy the subscription from them, I think.

Finance Assistant Workflow

And basically, we have multiple workflows. For example, this one is financial assistant. And it runs every midnight.

So workflow consists of multiple blocks, usually. And if you want to use AI, you also include AI blocks into your workflows.

So for example, this workflow runs every midnight. It finds new tasks in our task tracker in Project Pay, if you can see it. I'm not sure you can see it. Yeah.

It's like process some fields from these tasks, and then it run to AI blocks, which like your internal AI assistant, You track as our task tracker starts from scratch hit time, don't remember past actions, and so on and so on. So this agent is processing reimbursement and benefits requests from our employees, and it will

basically never update the status of the issue. But yeah, it doesn't work on Saturdays as humans. It's interesting.

Policy Checks and Human Focus

And this one will basically clear fields of this request from the human for, let's say, reimbursement. 1And this one will verify that it follows check fields, follow the policy. And also, if an employee forgot to attach something, it will write a comment to this task.

So it's kind of like simple workflow, which save. not time, but basically focus for humans. And humans can stop thinking about clarifying reimbursement requests with employees.

And it's quite customizable. So actually, our accountant built it so she doesn't want to do it manually. In any time you can build this kind of workflows by just creating blocks, drag dropping and so on.

Connecting AI Models via API

To use AI in this, so it's usually have some predefined flow. So it's more like a workflows and an agent, but it's workflow automation by using AI.

So to use AI, you usually need to connect to any popular models like OpenAI or cloud and so on using API.

Who knows what API is? Okay. So API is an interface for applications, software to speak to each other.

And if you want to connect your AI account from OpenAI to your workflow, you can create API key on the OpenAI side and edit here. So will call OpenAI on your behalf with some prompt and basically spend money on your credit card which is on file with OpenAI.

Other Simple Workflows

Yeah, this is one example of workflow. Let me show you something else.

This one is an active, but we had a GitHub outreach. For example, goes to GitHub, parses the GitHub, finds emails there, and stores it.

So there is actually no AI in this workflow other than this agent, which may do nothing. So this is just a simple workflow. It's not AI automation, but it's a simple workflow.

One more example, we had a birthday slack bot, which I wrote like five years ago, but at some point of time it broke. Admin manager like she didn't ask me to fix this, but she just wrote a workflow in an item which monitors birthdays and Congratulate people in slack.

So this kind of automation is easy to build by just drag and dropping blocks and It's also quite useful that you don't need to deploy it somewhere, because N8n is basically also a deployment platform. So you can create a workflow, schedule it, trigger it with a schedule, and it will work for you continuously.

Any questions about that? OK.

We have deadline reminders.

um like deep research partner so we had a slack bot which creates tasks if if we mention this bot and it's also implemented here we have some finance course I think we have some complex ones.

Product Digest: Aggregation and Summarization

Let's say product digest, which is basically, I will show you, but it looks scary. It aggregates information from Task Tracker Slack and other announcements and basically write the weekly digest for our team so we know what's going on in the company. gets lots of information, aggregates it, run AI to summarize and pick and prioritize and classify something that happened within the company, and post it to Slack at the end.

OK.

Recap of n8n

So this is NI10.

Introducing ONA

I also wanted to show you ONA. Disclaimer, this is our pet project, but I recommend you to use it. It's free and we have not implemented payment yet.

So if you will use it, we will lose money, but you will get value. So this is a more agency workflow automation, and I will show you how it works.

Natural-Language Automation

So this is, It's similar to N8n, but you don't need to understand what blocks in what order to build. So you can specify what you want to automate and your task in a more human language.

Connecting to Your Tools

And it connects to your... different tools. For example, this one is connected to my Gmail, Google Calendar, Google Drive, and so on.

And I can, for example, ask it to write a draft on my last email. Since it has tools, it will speak about MCPs as a tool. Basically, AI is using tools to connect to your systems.

It will get emails. So it will fetch emails. We'll probably find my last email, and we'll draft an answer to this email.

I hope it will work. Okay, yeah, it works.

So I've got some agenda from Paul about some meeting. Probably it was a newsletter, but anyway. And basically, yeah, this is this email.

Email Drafting Demo

And it's creating a draft there. So it's AI that understand what I want and operate within my tools.

And in my email, I've got, where is it? Where is the draft? I did not write a draft response.

So while it's doing so, I don't have much time.

Turn Conversations into Reusable Agents

So you can also, after doing some stuff, you can generate an agent based on your conversation, and it will repeat the same task again whenever you need it.

And let me finish this one.

Q&A Session

Maybe any questions about agents overall or anything I'm showing?

Yeah. What is the most useful agent you guys have built so far?

Oh, that's a good question.

Most Useful Agents Built

We built lots of agent around our HR process for communicating with candidates, some preliminary screening, and so on, customized for our guidelines and our use cases.

The useful one is, so we have team meetings and we record them with note takers, as probably everybody are doing nowadays. After the meeting, ONA takes a meeting transcript

Analyze what was said and do some actions so we can on a meeting we can say Like on I create a task about that for this person with a deadline for tomorrow And it will do it's like no human will need to do it. We just like Come on a meetings which are transcripted by AI and then processed by honor we just do like tell what should be done. Some easy like, um, coordination, like project management stuff.

Uh, and after the meeting it's done and uh, it writes to our slack, uh, transcript and I was like, like summary and uh, what have been like done, like what was like updated and the tasks and so on.

Yeah. So this one is, it's going to be, yeah.

Maintaining Context and Managing Cost

How do you maintain context within your agents and how, because there's a token limit regardless when you make calls, There is no like it's more like a technical question and there is no like easy answer like there are

So especially when we have lots of connected tools, AI have lots of options to do and sometimes use wrong tools or pull all emails. Yeah, once it pulls, like lots of emails from my account, and I could spend $10 on just one action.

Yes, it's an issue, but we try to improve problems. Yeah, there are lots of nuances of making it work.

Memory Strategies and Prompting Discipline

We have memory, which is like multilayered memory. long-term memory, short-term memory, and so on. So there are multiple hacks, engineering hacks around it.

And also, the important part is to keep agentic prompts, which we internally use. So it's mostly the question about internals, to be more focused on do this, but don't do everything, but still do this. So sometimes it's lazy to say, yeah, I can do it, but should I do it, and so on?

Yeah. Yeah.

But for example, this email draft. It's finally here, should be here. What's going on? Where is the draft?

Actually, we have draft, two drafts because I asked it twice. So it created a draft which will reply to this e-mail which should not be replied, but anyway, and I can generate agent. for example.

So based on the conversation, it extracts what was done, what kind of task I've done. And I can just run this agent or, for example, I can schedule it daily. So for example, every day without even my intervention, it will draft the answer to my email.

I can also ask it to draft all answers for past day emails and so on. So I can wake up and go to my Gmail and already have all pre-drafted.

Multi-Tool Workflows and Limitations

But the nice part is that it can connect to multiple tools. I'm going to connect it to Google Docs, to GitHub, to Slack. And it can pull some information.

So for example, if the email answer needs some information from your Google Drive agent, sometimes it's not smart enough to go there and look for this information. So it depends.

This one and similar systems, there are similar startups doing it. It's working for simple use cases usually when you do something with maybe a couple of tools, a couple of business tools like email and Slack, for example, or like Google Docs and JIRA and so on.

But models are getting smarter and smarter, and tool use is getting better and better.

Getting Started and Best Practices

1I recommend you to start using it and start using NA10 or ONA and you will have some workflows automated, and they will work better and better every month where you have AI part. So basically, you'll be able to focus on more interesting tasks and work instead of repeating the same every day.

Yeah, so this was about the agents and workflows. If you have any questions, yeah. Reply the message. Like text message.

No, I think no. There is no like, so it's much like online tools, like services, like stuff. But it's a good idea, yeah. Yeah, what's up?

Yeah, it's not yet. Yeah, yeah. Usually it depends on your prompts.

If you build a workflow on an agent, you can, for example, specify more details in this chat where I answered email. I can ask it, verify that it has a polite style or something else. The quality of output depends on the amount of tokens, energy, and money AI is putting into your task.

If something is critical, you can put more prompts, which will basically end up with more tokens spent and more money spent. But also, I recommend you to use AI automation right now for low stakes tasks.

For example, in marketing and sales people are using like a lot of AI voice to call potential customers who otherwise would get lost.

So for example, customers who already ignored all the calls, canceled subscription, and so on. So if you call them with AI, even if AI will make a mistake, it will not matter because this customer is already lost. But still, you can probably return 1% of the customers for very cheap, because AI voice is cheap nowadays.

So try to use AI where the cost of error is low. ask AI not to take a final action and have a human in the loop to verify this action.

Other questions?

Security and Privacy Considerations

Is there any security or privacy issue with this? I was thinking whether you have the memory to memorize some kind of important message. Yes.

Overall, there are privacy issues around AI. You can address these issues by controlling

On-Prem and Local Models

For example, like N8n, you can install it on your servers. If you have IT team, it can install it in your infrastructure. And for AI blocks, you can use local models, which are open source and can be installed in local infrastructure. In this case, your data doesn't leave your company infrastructure.

So it will be sealed in your company, or you can even run small models locally on a laptop. So potentially, even now, but especially in the future, we will have these kind of tools locally, and we will be able to run these kind of tools locally.

Local Assistants vs. Cloud Models

So your, for example, personal assistant can just sit locally, and it will use AI, which is also local. It will not be as smart as proprietary remote large models, but it will be smart enough to do most of the tasks, like write emails, schedule meetings, and so on.

But if you're working with some company and send the data there, you should trust them or read the privacy policy, this kind of stuff.

Sorry, just one more question. Yeah.

Merging Meeting Audio with Slides

So one thing is you mentioned previously in the conference, not conference, the meeting, the Zoom meeting, whatever, you take notes. That thing is, in the meeting, we have some slides, for example, something so. And then we have the talk or discuss.

How you can merge the information together to load the slides? We just use like third party note takers, like Fireflies, TLD. So there are like third party note takers.

I think they do not use video stream. I think they use only audio. Only audio? Yeah, I think so.

Yeah, but I'm not sure. But I think like they mostly use audio. Yeah, but probably they will evolve to use also video stream to incorporate into transcript and summary.

Yeah. Yeah. Yeah. Yeah.

Yeah. Okay. Yeah.

Who Can Build These Workflows?

You mentioned you had admins within your company that manages creating workflows. What is their tech literacy level? It's very low, very low.

So they're not developers. They're admin managers who used to do reimbursements or travel for employees or organizing stuff and helping others to be more productive.

So they're not developers. They don't have any tech literacy.

Nowadays, it's much easier to build stuff.

Defining Custom Metrics and Quality

when we have an agent like what he was talking about, so you need to figure out if the output is correct, right? And you want this to automatically happen, which means it updates its environment by having a feedback loop. But how do we define the custom metric for each agent? Can you give me an example of how you define a custom metric?

For example, sometimes like the easy way, please let me know if I should stop. I keep giving... Yeah, yeah. So you can provide some examples.

So for example, for like communicating with candidates or like screening, like HR screening. Basically, within a prompt, we provide some previous examples, like this was our communication and so on. It calls like few-shot prompting. Basically, you can provide some...

good examples of how you usually do this task and it works like most of prompting techniques are similar to techniques for like managing humans. If you have an intern and you want intern like to help you screen like resumes or communicate with customers, you are showing your good examples to this human intern, and then this intern will produce better results. The same for EA.

But yeah, we specify what is more important. And yeah, it's the same as for humans. So if you want to ensure the quality of human output for low cognition tasks.

So AI is kind of not very smart, but it can do low cognition tasks on scale and mostly free. So we do the same. We specify what's important, what to follow, some rules, and so on.

Keep in mind that if you specify too much, AI can basically get lost in your instructions. So try to narrow down your instructions and prompt to what's really important, and don't mention what's not important.

Thank you.

Conclusion

If you have more questions, I will be here.

Thank you.