Being able to run proactive personal agents that do things without your involvement is becoming so much more important now as models become so much more powerful. So I wanted to share with you an amazing release from deepseek that means where their harness will constantly watch a GitHub repo for issues that you may raise and it will just get in there and start developing it all on your machine and submit a pull request ready for you to merge. Or you can have it just automerge itself. But we will get into why you should not be doing that. Now, there are a few little bits and bobs in this setup that you should take note of. So, I do recommend watching this step. However, check down in the timestamps where you can just skip ahead of the installation just to see what this thing can do. This is basically what we can expect. We have a pull request here where I just asked it to create a green version of my personal website and it has created this really lovely pull request, very, very detailed with screenshots included. And I am really happy with this. This was all done by Deepseek V4.1 Flash, which low key is an amazing model. So, let us just get all of this set up and we will go through it together on how you can get this proactive agent model running. Now, it is worth saying that I have a Deepseek account with an API key. Honestly, it was set up a while ago. You might need to copy a Deepseek API key to to be able to use this with credit on it, of course. So this is the GitHub repo from Deepseek. Now they do explicitly say this is a research prototype for reference only and it is not an official Deep Seek product. It is open source so you can adapt it and they actually do recommend you adapting it to your needs as opposed to using it wholesale. However, I used it and it seems to work pretty darn well just on its own. But I guess they are just warning you that things are likely to change or or whatever. The idea of it is to take as little human attention as possible and give as much value as possible. It speaks only when someone must act or can use a result. So, it is just trying to get out your way and only involve you when you are absolutely needed. Now, it is a hefty setup process. So, what it literally asks you to do is ask get your agent to run this request. So, we will copy that into my clipboard. First of all, let us clone this repo. That is step 1 is actually cloning the repo. So inside of my sites or whether you want to download this sort of stuff, I am going to go get clone and then paste in that vaper. I have already done this, so I am going to rename this here. I am just going to call it deep seek agent, but you can just leave that as it is. CD into that and we are just going to run claude or whatever agent that you prefer. And I am literally just going to paste the read setup MD in this repository and set up DSH super agent with me. We will rush through these because there is a lot of questions and it will take about 5 mins of your time. Cool. Question 1. Should I continue the existing Super Agent profile and keep its configuration? If you say yes, I will keep the profile's current settings and follow the upgrade steps. So, nothing configured is either lost. If you would rather start clean, say fresh, and be aware that fresh setup replaces your profile and deletes the callers patch yl. We are going to encounter a lot of this stuff where it is we do not understand it to be honest. And I think that is part and parcel of this research preview here. The defaults as I set them up previously seem to work fine. And as I think once you start to understand it and start to get bit nuance with your needs and requirements, you can obviously just go in and edit this and figure it out later on. But we will try and get you set up as quickly as possible. Question 2 sources. Should the agent take requests from GitHub only or from FAU chats as well? Up to you. Face you. I I do not use FUEU, but I am assuming it is some sort of Git alternative. We are going to do GitHub only. So, I am going to save that cuz I will jump in here and say what it is basically doing. What we are going to be doing is setting up a GitHub profile just for this agent. And that is what the 1 you saw the comments for earlier. And it just means that you can track and understand the the things it is doing versus stuff that you are doing. and it is going to be working with a local folder structure, a different clone of the repo that you give it to it because remember this only works on 1 repo and it works in its own local folder on your machine so you can continue to work on your local folder on the machine. Question 3, repository. Which github.com repository should the agent work on? Now, I have already set it up as my personal website. Let us go with my podcast command AI here. So we will copy that and paste that in there. Question 4 agent account. Which GitHub account should the agent act as? Recommended a dedicated account which we are going to we are just going to set up for it. So dedicated what should the dedicated account login be? Recommending this. I have already set up this 1. So, we are just going to say command AI agent just to give it a different name, but you can name this whatever you want. So, it is saying it is available. So, we need to set up this account. Who may ask for work? Which GitHub login may ask the agent to do things. Now, I am a 1 man band, so I will just say me, but you will obviously put in the usernames of the accounts that can actually request work. So, if you are in a team, this will be the multiple different usernames of those people. I am just going to put myself. Question 6, check out which local clone should each ticket start from? Recommended a new clone at this. You already have a clone at site. So that is my 1. Let us go to this 1 here. Just paste that in there. Who should the agents commit be named as the author recommended this? So we are going to go with that recommendation. That makes perfect sense. And automerge. This is really important. Now, should the agent turn on automerge for its own pull request once they are ready and approved? Recommending no. So, I do agree with this. But if the more confidence you get, the more you understand what it does and how it does things and and whatever, you can turn this to yes. Now, while it is doing that, let us set up a new GitHub account. Now, GitHub being GitHub, I had issues with this, so just follow along carefully. Now, I would suggest doing this as an incognito window just because GitHub is GitHub. Making sure the username is the username we copied from earlier. Now you might see this which GitHub sort your out. This is mental that I cannot sign up for an account again. Now eventually you should get a code here which means the signup was successful. So again yeah we created the account and we need to run this to run the agent. So in here we should see an invite and here we go. And now this agent has access to the command AI repo. We are going to run this now. We can just do that there. Run the login script. Take the code. Authorize and we are all set. So now the super agent is running for command II show and we should see the panel here. So here it is. This is this is the harness. You are seeing some remnants of a previous run that I have done before here, but this should be blank in your instance. And this is a an issue that we have at the moment. So, Cabaza's information is not up to date. It is still from the template of the hosts page. If we look down here, there is just some dummy information. With this intact, now I can mention the command AI agent and say can we take some interim information from Kabaza's website to his mentorship page cuz this seems to have a bunch of stuff about him on there and leave that as a comment and in just a few seconds we should see it get picked up by the command AI agent. Now, if you are finding a lot of value in this sort of content, then I would really appreciate a like and subscribe at this point. Now, it is important to note that the fact that you can only assign it to one GitHub repo is a fundamental issue with this whole system. If you can get this running in an isolate environment on its own machine, then you will not have an issue theoretically, but the limitation is pretty bad that you can only have it on 1 machines. I encountered issues setting this up just for simply for this demo. being linking it to to several GitHub accounts running on the same machine and it looks like it picked that up. This is all using DeepSeek V4.1 Flash. This is an excellent model from my testing. It is working away here. It is not contrary to what the actual model name is. It I I found it is not particularly like lightning fast. However, given the price, given everything, it is still this is obviously something that is going to be working in the background and not necessarily something where if you need something done, you should probably just do that yourself. But this will just work in the background, get things done. It is also taking screenshots as you saw from the intro. It is a very, very thorough model. Now, you can go into the settings and models and set up your own provider or indeed your own model. So if you if you do have an open open router account or something like that, you can actually use DeepSeek through their servers as opposed to DeepSeek servers, but you will of course have to set up the API key and all the rest of it. But the harness itself is something completely separate from the runner. I think what this is actually doing, there is a runner that is constantly checking the GitHub profile with all the permissions and etc set up, which is then prompting the harness to actually action what it finds. So it looks like it is actually created a pull request here. This is an automated agent or human ready for review and about profile. So yeah, if we go to pull requests here, Phil Kabaz's host profile. And again, this is the this is this is what so amazing about this model. It is so considerate and so thorough. The why, the what, before and after. So we can actually just review it here. I am already seeing here the latest should probably be well we have got some dates there we need to maintain just to maintain consistency for the rest of the website but so I guess here we can actually just say maintain the dates next to each CV entry finalized on host as latest similar to Samuel. Whoops. We need to actually mention the agent. Now, what I have noticed it is doing is actually it is going into it is actually going into LinkedIn to find the dates of Kabaza's experience, which is really nice. Again, really thoughtful, religion good model. Cool. And we have an update. And you can see the before being here. and we have maintained the dates and the current. So that is probably something we need to fix too. Obviously could not find the end dates and that sort of stuff. So we will get Kabaza to update that. Now what I have done is actually prompted Claude to add Kabaza as an agent as someone who can trigger the agent. I will get him to correct the dates. But this is exactly what we want. As the agents get better, more capable, and more trustworthy, we are simply becoming these orchestrators, just popping plowing them to do the job, to leave you with a pull request that you give the thumbs up on or you know, leave comments or refine it. Now, I know a lot of you are not going to be there yet, but I honestly think it will not be long until we are there and we are getting it to do more elaborate, more things. Now, Deep Seek do say that you should do small requests here, and I I completely agree with that with we are changing very simple things here, but I hope this shows how capable these models are and how this tool can be a a stepping stone into being proactive about your codebase as opposed to reactive or being the 1 to actually be prompting the agent. So, let me know. Are you someone that already uses some proactive agent of some description? Do you enjoy Deepseek? Have you tried DeepS v4.1? What are your experiences with it? That will do it for this week. Like, subscribe if you have not already.