USEFUL Agentic Workflow: AUTO-Updating Blog with Claude 3.5 Sonnet

we started with prompts we then explored prompt chains now it's time to put it all together into an agentic workflow all of these prompts tools libraries AI agents and chatbots are completely worthless if they don't help you solve a real concrete problem let's look at one of the many unexplored use cases you can use a gentic workflows for to drive valuable end results so here's a simple blog post you can see we have a big H1 we have a couple PS we have a small list here and then we have a table but of course this is not a normal Blog the data that drives this table is built entirely off in a gentic workflow that updates by itself without intervention let's go ahead and nuke some data so I can show you exactly what this looks like so I'm going to remove the Claude 3.5 Sonet Row from this table and our agentic workflow will notice that there's a new model that hasn't been captured and it will read it so I'll move over to the data set and I'll remove clae 3.5 Sonet this is the Json row that is populating this table and now you can see we don't have that row it's completely missing so now we're in the state right before anthropic announced Claude 3.5 Sonet so now let's go ahead and run this agentic workflow and then we'll discuss exactly how it works so you can see it's opening up every page it's doing the scraping it's running the retrieve step it's running the trigger step it's running the agentic step and we're going to break these steps down in just a second including the notifications that you see on the screen that's my mobile device right here and that's going to be one of the critical steps of this agentic workflow we need to know when something important happens if we're by our primary device or if we're out we need to know what's going on with our workflows right now we are running our agentic these are our prompts our AI agents our flows you can see there we detected a change and now we have a brand new notification I can also open up these notifications on my device here and you can see here we have a new model Claud 3.5 Sonet has been added to anthropics lineup with an input cost of three an output cost of 15 so this is awesome right this happened automatically I was notified right on my device this is one of my favorite notification tools this is called ntfy also pronounced as just notifi and it's a really really clean simple HTTP based Pub sub notification service definitely check this out I'll leave a link in the description but I think this is one of the crucial steps of building out agentic workflows which we're going to talk about in just a second I think there's a flow to this I think there are specific steps that I keep seeing as I keep building out these workflows and in this video I want to share them with you and walk through this code a little bit to really show off how you can build agentic workflows end to end so let's go ahead and hop back to the site and actually see that change so you can see we now have that Claude 3.5 Sonet back in our table the data driving this table the data driveing this blog automatically updated it and it's all thanks to this full complete agentic workflow this is where the Channel's going this is a piece of the future that we have been gnawing at video after video building software that works for us while we sleep so let's go ahead and dive into this let's see how it works I manually kick this workflow off how this will actually work work is it'll be scheduled just like this and it'll run over and over and over so let's go ahead and just kick this off while we discuss this this is just going to be running in the background so imagine you have an agent farm that is just running arbitrary agents for you it has you know 5 10 15 different agentic workflows that operate on your behalf in the background while you're going about your day this is one of the key steps of the agentic workflow we're going to break down this is the trigger step so you can see here I'm using this simple python Library called schedule it's very very clean simple and easy to read you can see every 10 seconds I'm running a test job and test job is just going to print and every minute I'm going to run the real prompt chainable for llm prices this is the full workflow that we started building in the previous video you can see the workflow just got kicked off again it's been 1 minute so our agentic workflow is automatically kicking off it's running the triggered step it's running the retrieve step and now it's running our agentic this is where it's run running prompt chains on the scraped data and it's making the decision whether there are any new models on any one of our configured items and you can see here it decided that there are no new models detected there's no summary to write our workflow just continues to run in the background here I think breaking down your workflows into concrete steps is going to be the most helpful for you so you can see the first thing that happens is of course the workflow gets triggered next We Run The retrieve step this is where you're going to be pulling in all the information you need for your agentic workflow to run sometimes this is going to be from a database sometimes going to be from a file in my case I need to scrape a couple websites and then operate on the information scraped so this is my retrieve step I'm saving this information so that future steps can work on it so the next step is the agentic step this is when you're actually running your prompts your AI agents your assistants whatever combination of prompts you've built up this is where you're going to run that so in my case I'm going ahead and looping through each one of the providers and this represents a single block here so we're going to Loop through each provider we're going to look at the scraped content for that specific provider so what was the content scraped for instance from anthropic and then for each one of those we are going to run this prompt chain we covered this in the previous video when to use prompt chains I'm going to link that in the description definitely check that out there's a lot of valuable content in that video including this minable chainable class that you can use to build out promp chain CHS that have back references to the outputs of previous prompts and you can also pass in context to use in every specific prompt again I'll link that in the description there's a lot of value packed into that last video so we run our prompt chain now after this runs we're going to have our table results for this specific provider we then store those results and that completes our hentic step after that we now are running the act and notify step if you are running tool from your agents your act and agentic step will be intertwined you can also notify directly from your agents I like to separate these steps out I'm just running prompts and then after that I'll operate on the results and now I'm looking for changes detected so I have this function that detects if something changed between the new scraped values that you can see are getting scraped right there from open Ai and than the previous results that we already have saved right because that's how we can tell if something is changed right we it against the previous result you can use an llm for this you can use raw code for this that's completely up to you and your implementation I do think over the Long Haul more and more code and we'll talk about this when we go over the five tips for building agentic workflows I think a trend we're going to see is as these models progress and as we improve our abilities to operate with models and prompts and prompt chains all the way up the stack we're going to be writing less code and handing off more of the work that needs to be done to agents so we get changes detected this is is going to be a simple bulling if it's false this is just done right and we've seen this happen several times here uh no summary detected that happens if there's no summary detected here by our agents we summarize the changes if some changes were detected using the incredible Sonic 3.5 model here just to look to see if there are any changes that are significant if there are we go ahead and update our table rows and then this is really cool this is where our notify and our commit and push step happens if we actually detect changes and we can dig into this prompt this is not something we talked about in the previous video so this prompt is pretty simple basically you give it the previous Json data current Json data and then we just want a summary of the changes in this specific Json format right so we're looking for this we want a summary if there is one if there's no significant change just return you know true or false and you know again we're running a prompt that drives the outcome of the application that's what makes this in a IC workflow it's driven by the reasoning of your prompts after we have our change summary we then check that variable that's returned again right from our agent and if there's a summary detected we then run our notification right that's how we got our notification to our mobile device and any device we have subscribed to this specific notification topic that's when we got our message there and then this step is not filled in yet because this code isn't live yet but we have this commit and push step step this is currently unimplemented but you get the idea right once our gentic workflow detects a change and we push to this file with the change inside of it we are going to commit and push this to production or you know staging however your workflow works we are giving our agentic workflow full control and we're giving it the ability to say when our agent says we have changes we go ahead and push those changes up we let our agent drive that behavior and then that completes the final step of course you can see there we're kicking off again so that's a large part of how this works let me just go ahead and work through those individual steps one more time I think that's the most important thing that can really help you build and drive your own agentic workflows this is a sequential agentic workflow so there are predetermined steps that I'm driving my prompts through that I'm driving the prompt chain through this isn't a you know group chat conversation or a multi-agent conversation the common example you've seen is do some research on this topic and you have you know four agents with different capabilities and they just run off this is what I like to call a sequential agentic workflow so we have predetermined steps for how exactly we want things to work and what exactly we want to have happen right so I think they're going to be different types of agentic workflows all working up to the final type which is a workflow that is basically prompting you for information and then it goes off and does all the work that's for a much much uh later video let's go ahead and look at these steps so we have the trigger step this starts your agentic workflow you hit an API where your scheduler kicks off like we're doing here right we have auler that's triggering every n amount of time to start the workflow a lot of these workflows are going to be triggered via a you know UI so that's the most important part right obviously uh you need to start the workflow somehow so that's step one I think step two is then retrieve in this example we are scraping a couple websites to gather information about the models that we want to display on the site right so you know we're looking at U vertex we're looking at open ey we're looking at anthropic just pulling in some information letting the agents run on that and then that all gets formatted into the output here right but we first need to retrieve information this will sometimes come from an API a database you know a local file uh message you know some type of Pub sub or web scraping in this example right so this is a ret step I know I've mentioned that I am anti- rag on the channel before that is true but for sure there are some cases where you just have so much information that you need to use some type of chunking mechanism my advice there is to keep the chunking as simple as possible as the context Windows of every model increases and the price goes to zero and then the next step right so of course this is what really differentiates agentic workflows from just normal code right because at the end of the day that that's exactly what this is right it's code code powered by a reasoning system and so what we end up doing here we run our prompts that solve your problem so this is you know your prompts prompt chains or AI agents if you're using the assistance API you're doing that this is where the tool calling happens this is where all the magic happens right then we have the ACT step this is where we actually act on the information retrieved calling tools and this officially completes the task at hand right so now we're updating files we're updating databases we're updating the API we're updating the UI Etc right in our case we are updating this file here right this is the end result this is the action that is caused by the workflow so we're getting these clean rows for our Blog then we have a learn step so I'm not going to dig into this too much here um this is a topic for another video this is where your prompts and your agentic step this is where it learns so my workflow does not have a learn step this is where your agent self-improved we're going to have several videos on this uh we're going to go ahead and just kind of gloss over this step for now because this is pretty advanced stuff that uh I'm still exploring and many engineers and researchers are still exploring uh so then we have the notify step right so this is the notification you saw I think it's really important to be plugged into your agentic workflow when something happens you want to know what's happening and you want to be able to observe when it's doing the right thing when it's doing the wrong thing so that you can come in improve your workflow or teach your workflow how to improve you just need to be aware of the results of the system just like you would any system nothing special there so that's the steps for the sequential workflow so that's what that looks like right let me open this up so you first trigger your workflow to get it kicked off you retrieve any information you need to run your prompts your agents Etc you then run your agentic this is the key step you then act on that information you act on the results of your agents you then learn if there's anything that your agents need to learn from that Run Cycle from that iteration and then you notify right this is where you're alerting yourself this is where you're making sure you're logging and making sure that this information of this agentic step can be easily attainable for this workflow I built up this simple system where I have the outputs running and it just uses you know a date time stamp and it contains all the information I need to look back at the individual runs so here we can see the model pricing output for this one even if it wasn't used and pushed to production we can also see the scrape results we can see the prompts that were being built for each provider we looked at a lot of this in the previous video again refer to that to dig into more detail of what exactly rest stored here but I think it's really important to have a full built out logging system you can dig into I'm using some really basic file logging here because it gets the job done fast so I think this is a good breakdown of the sequential agentic workflow steps right trigger retrieve run your agentic act on the results of your agentic update your AIS with any important information that will help them improve then you notify yourself you notify the Builder you notify and you log the results so last thing I want to talk about in this video I want to give you five quick tips for building out these agenic workflows the first thing that comes to mind is be incremental start with the bare minimum prompts start with the fewest abstractions the fewest layers and then slowly give your agents more abilities and autonomy slowly build up the prompt chain slowly make it more complicated give your agents more tools over time really really start simple this is the fastest way to get some Real Results second tip I mentioned this before but use great logging and notification I think it's really really important to have a tool like notify that you can you know easily alert yourself easily alert you know any other Engineers that you're working with and and just really keep track of your system because how will you improve your system if you don't know the outputs and you can't debug it right so great monitoring leads to Great debugging and system level understanding uh don't skip this I I highly recommend running the state-of-the-art models first uh don't try to Chimp out on anything that um isn't just the best when you're building out these agent workflows if you're really trying to drive valuable business results just start with the best and if you can then slowly replace your models with cheaper models over the long term I've mentioned this before on the channel but all the costs of these prompts of all the great high class models the cost of all this stuff is going to zero it's all getting commoditized so you want to be using the best tool for the job and just getting used to the capabilities available to you right uh llms are just functions it's important to engineer your AG gentics as you would any other function right when it's time to call your AG gentics when it's time to test when it's time to maintain these systems it's just a string in string out function I think as epic and as awesome as these tools are I think it's important to keep that in mind right underneath all of it you have a black box where a string goes in and a string goes out all right so build for tomorrow this kind of Loops fully back to tip one be incremental over time you should be giving your agents more abilities and autonomy and writing less code I do think that in the end you'll find yourself writing less code and giving more tasks and work to your agents I think this will take time I think we're going to have to build a ton of agentic workflows which is exactly what we're going to be doing on this channel but I think we will get there over time we're going to be writing less code we're going to be doing less of the individual lines and instead of using you know conditionals and if statements and loops and all this stuff we're going to give those capabilities to our agents and they're going to do it for us so that's where we're at let me know what you think about my five tips here for building agentic workflows do you agree do you disagree with any of these you know let me know what you think about the sequential agentic workflow steps these are getting built out more and more not a ton of people are building out full end to-end agentic workflows this is where we're going on the channel this is what it's always been about there are a lot of cool stop gaps on the way but to me this is the killer use case for this incredible technology we're building out living software we're writing and Building Things that builds on on our behalf while we sleep and the auto updating Blog the auto updating table here is one of many first examples and pieces of living software we can build that uh we're just getting started with so you can see there uh you know still auto updating thanks to the Schuler that's it that's a Full Circle here and uh thanks so much for watching sub comment like and I will see you in the next one

Channel: IndyDevDan

Views: 4,793

Keywords: agentic workflow, agentic, agentics, prompt chain, prompt, claude, sonnet 3.5, anthropic, prompt engineer, auto updating, self updating, llm blog, llm blogging, llm seo, chatgpt seo, ai agent, agentic engineer, llm library, llm abstraction, framework, claude 3.5

Id: F0eOYrA6ggY

Length: 18min 27sec (1107 seconds)

Published: Mon Jul 01 2024