Summary
Notes
Transcript
as well I started this session unofficially sponsored by ClawStation which didn't work after my first attempt I'm now working on my goal to talk about learn about how an agent works in agent force Following the content from the Course on Udemy by Mike Wheeler. I love this name. Why I love this name, I don't know. I remembered, it's like I remember this name by heart now. Remembering names is a stage where it's like More important today So when we are not able to recall many names today, there are very limited number of names we can recall, right?
And that's the partOh, it'll... Being a human Okay. So this content from this course is like already outdated but it helps. Behind the scene, that will serve you well. so i will describe this Next first behind the scenes agent works on movie With the LLM to dynamically select which topics and actions to use. Behind the scenes. the agent works with the llm to dynamically select which topics and actions to choose Agent is An agent is built on three pillars.
Actions. instructions and Instructions, actions and I'll just recall from this process. There's a process to everything, right?
on the exam. And so the main points or takeaways here is that behind the scenes, the agents work with the large language model to dynamically select It has given me the third one, right?
Instructions, topics and actions. So which one comes over? So I have not actually or I have designed some picture in my head like how these topics, instructions and actions work. Actually we need to design, so we'll just now Come back to our real agents. Showcase time! My water bottle is now finished. This is the library there. agent for studio this is what it was called it Mm-hmm. This is my current agent on my website, three mistakes of my life.
and does it work see it doesn't work right now It won't respond to my hello. So there are a set of instructions, actions and Topics already, right? Ta-da! and that's the like core of every agent on Salesforce the topic like Are we on track? How an agent works in agent force? That's the question we are answering. An agent works based on three core principles, pillars. Should I call it a principle? No.
It's an architecture, right? Hmm. Oh, I did not open this agent yet. Should I have been more prepared to showcase the agent? this is not not not responding so I will actually just enable the previous version of my Agent. The Agenda is to understand how an agent works in depth. Right, do I understand? So I will try to Ask the question myself like do I really understand the topic. And we will ask this.
from ai assistant yay we'll generate some questions with the self right So will AI to use AI? Will AI to use to learn to use AI? I don't know what I'm doing. Okay. Yay, oh shit. Why is it like this is the older version of agent force. agent builder agent yeah agent builder and let me try to open the new one because what happened is the new one is that these are the two versions of my agent And this is the first time this Sidebar is collapsible in agent.
Salesforce platform oh my god I need to do it like thisEase out, I'm working on, I'm in a state of flow. I'm working on a state of flow. Yeah. It helps. Live broadcast helps. So these are the set of instructions, topics and... actions within Agent force configuration, my agent, three mistakes of my life, right? So how does an agent work? Again, look at the description from Salesforce. Agent works with the LLM to dynamically select.
just the LMK select actually not just so we need to understand that of the statement right Because there are many topics, because there are many instructions, because there are many So an LLM's thinking power needs to be used. Thinking, should we call it thinking? Or reasoning, we should call it reasoning. Reasoning power should be used to need to be used. to dynamically select the topics, actions.
So we need to be very clear about our instructions. the more clear the instructions and that's where we fail as humans what happened like I'll not jump many loops here The journey begins when an agent matches an event to a topic based on the that topics description. Read it again. The journey begins when agent matches an event. to a topic based on that topic's description. Then an agent of light, matches And even an agent three mistakes of my life matches an event Event what event now the event is like starting the conversation.
That's one of the events. So I think it should have been more general. But, let's see, matches an event to a topic based on that topic's description. So it's like, again, why is it saying topic? Let's ask AI. the instructions is No, let's... three Set the prompt. The content says the journey begins when an agent matches an event to a topic. based on That topics description now I'm confused about the event like what is the event the author is talking about here Yay!
We are using AI to understand AI. content from AI let's see what it says oh there's too much text Text is like-the difficult part In the context of the course, the event refers to a user interaction or a trigger that the agent needs to respond. So there can be two kind of events. It says interaction or a trigger. That the agent needs to respond to. So now it is very clear, right? The user interaction or a trigger.
an event is basically an interaction or a trigger that you need to respond to. Right, just to so my agent is not responding. And I'll just enable my agent first. I know my Tent. version of this agent Works as expected to some extent, right? So that is the current version of my agent that works. Oh, that is the current version of my agent that works 10th version 10th, right? And versioning is like the best part of having a software, building a software.
We can go back to the previous version and start again. That's what mobile Video games taught us, save the state of the game so that you can come back to it and start again. There are unlimited lives. You can keep coming back to the... working version right so now I'll activate this version this is what I know that it was. And this version of the instructions, topics and actions what does this agent do it responds to my the description of my project it wants it It describes my project.
So I'm data library. So I've connected this agent with the data library. So I'll refresh this page now. My agent was not working and this was, it was stuck at transferring. I don't understand this transferring keyword here. It shouldn't have been ideally like this. Why would it say transferring? It's not responding since like very long time. There's a broken funnelOkay. my pages are refreshing The topic, the question I'm answering is how an agent works.
So my this existing agent, my built, my Version of an agent. One version of this agent, right?
Working version 10 So this is like notion page.
It is like very stupid page. It is not accessible to I need to close this conversation now. Oh no, I refreshed the page. Should it have been there like this or should it have... Crash. It's saved state, so it's I think great to have this thing that my agent can continue the conversation even if my Pages refreshed. Yay! Do I understand the depth of my this thing is that a default behavior maybe it is because I have not customized my agent until that depth so now my agent responded I changed the version what is missing is that this icon should have been my very personal It should have been very personal message.
I tried to customize this it is not very good message Maybe It needs to be customized more. Okay, it could be done from somewhere in this part. The instructions, let's look at the instructions. Where are the instructions?
Ta-da! Oh, these are the instructions. Just for the topic.
Okay, Eats. Topic has their own set of instructions. Makes sense? Because ok that's not very obvious Each topic should have its own set of instructions is not pretty obvious. What I am looking at is like when the agent fires. Great. Conclusion starts Is it still in the set? That state is not responding. why was it like this like transferring There's one bug. let the notion agent Log the bugs, right?
Okay, and I will have I should log this bug in my Gate repo message to self hey thanks for making notes automatically dear notion I think I will have to pay for it sometime soon or I will have to find someone who will guard free. Ok, now my agent is responding. 10th version is responding. Now let's go back to the original question we were answering. Oh, I hit the wrong page. Okay. Journey begins when an agent matches an event.
Event could be firing an agent, like when I fired this, when I launched this agent on this page, there was an event. That was one of the events that took place, that agent was fired. Hmm, make sense.
Now, to a topic based on that topic's description.
I think it is also versioned I cannot change this part without creating a new version of I am resistant to the like. First instruction that was followed by the agent. That I am an AI assistant, I can help you. Thank you. so the instruction was Tell your agent how to greet the end user. The welcome message is sent at the beginning of each conversation. This is the instruction, I think. Like, there's a system instruction, I would say.
and you are so this is the instruction to the Soul You are a service agent that answers questions strictly from the documentation. Uh... this is the has i think that he's like limitations it is not really carefully defined let's try to use AI to describe this instruction in better with better clarity. I should also delete the version of the agent that is not required right now. So it is like cleaning up.
the agent level instruction is very basic. should we customize it for better functioning Did I ask the right question? Did we... Are we like dropping the right prompt? Let's look at this definition. Provide your agent with guidance on how to interact in all situations. So this is the like Agenda. What is the soul? This automation supposed to do. So my instructions Like as a guidance? It's like a .md file for but I have not done it yet.
When I asked this question, One of the times I did it, the agent force actually customized some instructions and said, you know, accept the changes. I just asked a question to understand something. It went ahead and made some changes. Not expected. Okay, now let's read it.
so a lot of things are missed in this Thank you. System instruction a lot of things felt missing So I should now. I felt I should expand and I asked the agent force within my agent force builder. To help me understand, right the limitations and recommendation and it said yes use these instructions for universal behavior what kind of universal behavior we are talking about Scope, safety, tone, refusal policy and fallback when info is missing.
Everything is missing in this instruction. Keep them short and stable. Avoid topic specific logic. Variables, tools or examples.
In charge bidding. And it crashed my agent, so version 13 is crap. Now I'm working on version 14 version 13 I will delete we are now customizing so I'm not Jumping to the basics again, ABCs. Why would it fail? Have I done the basics like foundation work very well I am starting from? the content of the certification course ABCs. Like very raw, very stupid, very rare, very... very soul oriented, very grounded, very clear I think I also need to take a break.
ok limitations low granularity. not enforce deterministic per task procedure. So it should not be about like per task. Right. These instructions should not be about partner. Shouldn't reference actions topic variables or transform. Why I'm talking about limitations is that I need to understand the boundaries I need to create. I need to understand what is that my agent should not be doing and what it should be doing will be more clear now in that way.
This is another philosophy that a chat GPT has like maybe slowly imbibing in us as humans that we should know what it's not When we ask questions, what happens in a conversation with chat GPT is that it describes. Um... What it is not. Whatever it's saying, what it is not, even clarifies what it is saying. That's the ideal way of having a conversation, doing things. And a conversation is a very scientific event that leads to maximum benefits for all the involved parties.
Well And every conversation is a funnel. every human is always in a conversation with themselves why I'm repeating that all in a very speedy way I'm recalling my own thoughts posts and Perspective about conversations.
With the Gemini.
create a quick banner Very Low UX. just text maybe with Scheduled break. for five minutes.
I had an expectation to make an image. Create image.
So I'm back What did we learn today? Hmm? Let me just... I'm going to connect my screen.
about open Claw? Station.ai but it didn't work. Let's see if it works now.
This platform, right? OpenClaw. No, clause station. Okay. Which is not working So just describing about the code what I understand is that it works in a way there are instructions there are actions and there are topics there are topics instructions and action but there are global instructions as well it starts with global instructions so my message goes. So then I figured that my agent doesn't really have meaningful instructions as agent level instructions, like the broader picture.
also be doing some kind of negative testing as well. I can answer questions related to the project documentation frequently ask questions about projects which projects help with Project Specific Issues What would you like to know about your project? Which project are you talking about? So I am also using this whisper flow.
If you have specific project or topic. So this is like bad. Bad response topic you would like to discuss please let me know. What is the name of the project I am working on? It's a wrong question. Any visitor trying to understand this project will not ask this question, right? I do not have any specific Any good? name of your project unless you provide it, please provide Okay, these are the like, stupid.
Responses of my agent. Yay, I found my agent is stupid. So it's like a-Revelation, my agent is stupid. I would like to say it out loud. My agent on my website, three mistakes of my life is stupid. Much like me. Okay, so now our response is yes, Drop a set of clear sections start with exemptions It said, great, I want to have full screen but I think I cannot Before drafting here, are some assumptions A few questions.
It's a great community with about 75,000 people. but not many are active less than 1% are creators. So yeah It's like the greatest opportunity of our time. Okay. The response was-Here is a concise way to define the audience in your agent system instructions for three mistakes of my life. Audience, you serve general website visitors and prospective participants Instead interested in learning about the project.
makes sense versioning is nothing But what all we would customize in the next version? There should be a checklist.
Wait it is not complete. we need to define the way the system will take care of hallucinations Now I am being very careful at defining the system level instructions Okay, so it's yet. Yes I can add precise policies to your system agent instruction. Before I draft them, a couple of quick questions. Source of truth.
To be taken care of by the builder that is me. In an ideal case, in the current state the agent should Provide the level of confidence. it has in the information As a disclaimer. Hm? and the confidence threshold is 75%. So does that. I am worried. I am saying 75% confidence. Should be the threshold. and it should state the confidence in each message. So when it is like 75% confident, about being 75% confident Would it contradict somewhere?
This is important. Why I am celebrating? This is kind of the first principle of doing things right. We need to understand the things we are doing. We'll go into GPDs. Where's my GPD? Where's my custom GPD? agent force architect this is my custom gpt million dollar consistency partner i'll just open it from here It's a custom GPD which is like which I built to keep myself on track. and It doesn't work maybe as expected so maybe this is like the phase of evolution.
10 minutes this is 7 about 7 and I will come back Yes! Go down and call me back. Come and tell me, papa. It's been 10 minutes. Thank you.
This is how I am training my child. Do it.
I don't trust it can create end to end action for me it will just create a dummy action. audience general public site visitors seek instructions about the three mistakes service scope, provide high level descriptions. I think I should copy these instructions and validate with the help of this custom agent.
Senseful, Because That will try to take care of the exceptions or The format, the agent force is not able to think, right? I want to use multiple Perspectives, skills, powers of artificial intelligence. Okay. Sensitive topics to refuse medical advice, legal advice. Okay. personal counseling financial investment I mean these are two vague topics, two general topics but let's see. Keep it. maybe we can review these instructions with some experts in the industry later yay Maybe I should give an instruction to the notion.
we will test this first in the Preview mode if they seem to be working like Yes, I will publish this on live broadcast on live website.
Debut changes project FAQ has been updated okay But why the system instructions are not updated? That was the goal. Asian forces to bed! Agent forces to open! Agent forces to open! I said it. Amen. It will be like super power in near future but it's not That is not... working as expected i expected it to update my system instructions and it is doing something like running very Very random. Okay, okay so I need to be very careful about accepting these changes.
Can I go back? Oh, something went wrong.
this is what The AI should not be doing it. Oh. and now the flow is broken Whoa, I broke Agent Force. Can you now specify what like what is that I am looking at Something went wrong. I should be actually asking questions Let's close the unwanted tabs. now look at the present faq okay It is not able to load. I will decline the changes So now some the Like two dummy topics are created and it is very I hope My agent still works.
Changes were not actually committed to the script. In preview I can prep propose and simulate but edit require an explicit script update if you do not run an actual script the system is just remained unchanged I'm going to stop I'm going to stop for today. So this process actually did not help me much. can it revert any fpq edits Yes Yes, revert the previous.
So I'll stop there. session for today I will come back tomorrow with maybe more planned process of doing this and we will customize the system instructions only currently it is like very blank Oh, we tried. Whisper flow is also failing. So my agents are failing. They are not really. Oh shit.
Revert the previous accepted changes and Update Dude, you are killing me. Yes, revert the previous account. rework on the system level instructions. Okay. Last 2 minutes for the day. it has been a fruitful learning building session and I will do it more often thank you Let us look at this once if it seems meaningful or not Tomorrow will probably start from scratch. but I will copy these Drafted system level instructions to my Notion page.
to be used again tomorrow. And it is again taking long Huh. should not have taken so long.
Come on The things are deleted. Yes. reverted reverted reverted reverted yes one section has been added to the system institutions. location login no external links anything and please no topic in these policies unless they That's very minimal But except let's just except something.
Try to be. Try to feel productive in the session, right? Okay, I'm not... activating this version yet I am NOT testing it for today but I need to save the state reset the simulator Hello.
support topic Services today, this is like very bad agent response, first response, so I will Say thank you for the day Oh my god, I thought I was sharing the screen!