Harness² : Google Open Sources for Self Improving AI Agents
Publicado no IA Dots ·
Agentes de IAAI SignalAssistente pessoaldot-ai_agentsdot-artificial_intelligencedot-googleidioma-enInglêsyt-ADbSlm4XsBA
Vídeo em Inglês, publicado pelo canal Codedigipt.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
The video discusses Harness 2, an open-source framework by Google designed to allow AI agents to automatically refine their surrounding systems and improve capabilities without altering underlying model weights. It explains the core concepts such as the propose-probe-compose loop, task-level versus harnessing-level recursion, and parallel versus sequential execution modes, while demonstrating a practical implementation inside an open-code environment.
Key points
- Harness 2 is an open-source framework created by Google to let AI agents automatically refine their surrounding systems without modifying underlying model weights.
- The framework exposes eight main agent components, including system prompts, guardrails, memory, tools, sub-agents, skills, scripts, and plugins.
- It operates using a propose, probe, and compose methodology where modifications are proposed, tested through behaviors, and compared to build better harnesses.
- The system utilizes task-level recursion for immediate task refinement and harnessing-level recursion to store learnings as persistent agentic improvements.
- It supports both parallel recursion, which tests multiple paths simultaneously, and sequential recursion, which tests paths one after another.
- The presenter demonstrates installing Harness 2 via an AI agent and running a file-search test to compare results with and without the harness tool.
Topics
Harness 2Google Open SourceAI AgentsRecursive ImprovementPropose Probe ComposeAgent ComponentsParallel and Sequential Recursion
Video transcript
Transcript from available captions; it may contain errors.
Hey, hi guys. Welcome back to another new exciting video, another new great open source GitHub repository, and its name is Harness 2, and it is from the Google. And they have published it as the open source because they want all of the AI agents that exist like open code, cloud code, codex, they can improve itself. So, that is why they are calling it as the self-improving AI agent framework. So, Harness 2 is an open source framework that lets the AI agents automatically refine their surrounding systems. I have tried this Harness 2 inside the open code. Here you see the Harness 2 loop that verified my experiment, and also this propose probe compose. I will explain you in detail what are these. Basically, these are the methods that they are using inside this Harness 2, and I am using this GLM 5.3 flash free model, and the max effort, and the testing that I have done it actually passed, and it actually working, right? Here you see that Harness 2 improved the harness and completed that task. And that they have used this K2. What is this K2? All of these things I will discuss with you in detail. But before that, if you are new to this channel, don't forget to subscribe this channel to get all of the latest AI content daily. First of all, they have published this Harness 2 as the open source for automatically improving the way AI agents work without changing the underlying model weights. Remember these things. This is the main important line. So, inside all of these AI agent, let's say inside open code, currently I am using this GLM 5.3 flash. Okay. Now, you will use the GPT-6 Astra or Claude 5.1 or Mythos or Opus 5.5 to get the best output, right? Now, the thing is Google is saying that you can use the best quality models, but at the same time, if you are using any medium quality model, and if you can improve the agent capability, then probably you don't need to use the best quality model. Means they are saying that you are using this GLM 5.3 flash, it is fine. But if you can improve the existing AI agent capability using this harness too, then probably you don't have to go with the another best model. Okay, keep the model as it is same, but just improve the capability of your AI agent. Currently I'm using this open code. So my aim is that I have to improve the capability of open code using this harness too. Okay. So how I can do that? So here you see that this is the harness too, and if you scroll down below they have mentioned one is task level recursion, another one is the harness level recursion. So what are these recursions? But before that you have to just understand the working architecture of this harness too. Okay, so for that here you see uh you have to just follow this portion. For that you have to understand one concept which is the propose probe compose. Now how you can find that? Just search here the propose and here you will find this. Okay, propose probe compose. What is the meaning? Propose is the uh is the way the system creates different modifications to the agent harness. Like I'm using this open code. So this propose step, this propose step is actually do the different kind of modification to this agent harness. Okay. And now next is the probe. What probe will do? Those versions actually perform the task, so their behavior can be observed. Means in propose step all of the modifications that it did, probe probe will actually uh test that behaviors. Now what compose step will do? The system then compares the resulting behavior and builds a better harness. Compose step will actually check that after applying in propose uh is the system got improved or is the agent or does the agent agent capability got improved? If improved, then it will apply that. Okay? So, this is actually the proposed probe concept. And now, as I told you that there is a two things. One is this. Let me show you again. Task level recursion and harnessing level recursion. So, what are that? But before that, let me tell you one thing. What this harness tool actually doing? Okay. So, it actually exposes eight component. Here you see that inside the AI agent in open code or cloud code or code X, there are different kinds of components are there like system prompt, guardrails, memory, tools, sub agent, skills, script, and plugin, right? Now, this harness tool actually exposes eight components. So, what are that eight components actually inside AI agents? These components are system prompt, guardrails, memory, tools, sub agents, skill, scripts, plugin. These are the kind of components that is present in any of the AI agent. Okay. Now, again I'm reading this line. Harness tool is an open source framework that let the AI agents automatically refine their surrounding systems like prompt, tool, memory, etc., etc. that components through a recursive process of proposing testing composing improvements. You remember that I talked about that proposed probe system? So, this is the way they are actually improving the agent harness. Okay. Now, uh this two things. Task level recursion and harnessing level recursion. What are these two? See. Task level recursion means let's say you have given any task to this AI agent and then using this proposed probe compose step, it actually refine itself and give you the best um output. Okay. So, here do you see? It actually followed this K equals to three steps and after that here is a K equals to two. So, it compared the K one step with the K two step, and then it give you the best result. Okay. So, this is a task level recursion. And after it will give you the best output result, it learns, okay, that if I follow this step, then I can get the best result. Okay. Okay. So, then it will update as a persistent improvement harness inside its agentic capability, and that is called harnessing level recursion. So, first it will do the task level recursion, and after that whatever it will learn, it will store that as a harnessing level recursion inside the AI agent. Okay. I hope you got the point. Now, here if you scroll down below, then you will find this installation procedure. But, don't worry. I will show you a simple installation procedure. Okay. But, before that, let me show you another two term. One is this parallel recursion, and another one is the sequential recursion. So, parallel recursion means, let's say you have given any task to this AI agent, and that task can be done in many ways. Okay. But, let's say if you follow the step one, then or let's say the path one, then it will take some time. And if you follow the path two, then it will take some less time. But, if you follow the path three, then it will take some very less very very less time. But, path two is the optimized way. Okay. Because it will give you the high accurate result. So, in parallel recursion, it will actually try all of these all of these paths in one go. Okay. So, how do you see? It will actually try all of these paths one go. But, in case of sequential recursion, what it will do? First, it will do the path one, and then it will actually try the path two, and then it will try the path three. Okay. So, in sequential means, sequentially it will check all of the paths. But, in parallel means, it will parallelly check all of the all of the possibilities. Now in Harness 2, and this parallel is the by default capability they have attached. But if you want the mode as a sequential, then you have to explicitly mention that go with the sequential steps. Okay. These are the successive steps. Now in my case, here you see that it actually followed the successive sequential steps. So that's why here you see that mode is mode is sequential, and it actually tried the second path after the first path, and then it is asking me that do you want to go with the third path or not? Okay. And I also recommend you to follow this sequential mode. It will actually help you to better identify. Okay. Now actually how to use it? Let me just show you. So inside this open code, here you see that if I just go to the top, here you see that I just asked it that see this repo Harness. I actually gave this repo link, and I asked it explain me. So it explained me, and it has also got the context of this repo that what it actually does. Now after that, I asked it that install this Harness 2 inside here, this machine, and repo link is this. So it actually installed that Harness 2. It actually found the version that is required to install. It got matched with the environment, and after that it modified some of the open code files. Automatically it did. I have not done anything for this setup. Okay. So that is why I'm recommending you you're recommending you also do this setup like this so that you don't have to manually execute any of the commands they have written inside their installation procedure. Okay. It will be very easy for you. Just give your repo link to this agent, and it will do all of that things. Now here you see Harness 2 is installed and fully working. And after that what it does that um Sorry, after that I actually asked it to test it. So, for that uh let me show you that uh comment. Let me show you that comment. Here you see, I asked it to create a folder with few text files inside the notes folder. Create this apple.text banana.text and orange.text. And inside all of these text files uh write some of the lines. And after that, the agent task is find all of the files inside this folder that contains the word AI and tell me their file names. Now, if I show you, here you see that inside my folder it actually created this notes and this these are the text file apple, banana, orange. And inside this text file it actually have written some of the text which contains some words. And I asked it to find the files which contains the word AI. And when I gave this requirement, I first use it without the harness tool, okay? And without harness tool it actually uh found this file name successfully. Now, again, I actually performed the same task. Now, this time I asked it to use the harness tool and improve the harness and then complete the same task. Now, it actually used this proposer this that probe and that concept, okay? And after that here you see that uh it actually did many task. Okay. So, I actually if I go below, here you see let me show you. After this K2, means K2 means uh it actually comparing the K1 step answer with the K2 and it found that K2 is the another improvement in the harness and that is why they have written harness tool um uh improve the harness and completed your task. Okay. And uh here you see that uh this is the answer from the improved harness apple.text, orange.text. And uh here you see, these are the result that I have found. Okay. So, that means in this way you can actually improve the harness capability. So, this is the thing, guys. And uh I hope you have got the point about this harness, too. And uh if you have any question, just let me know in the comment section. And another thing is that uh if you are visiting this channel for the first time, don't forget to subscribe this channel. Don't forget to like this video, also. See you guys in the next video. Thanks for watching. Bye-bye. Take care. And uh also, don't forget to uh like this video. Don't forget to subscribe this channel. And watch the other videos, also. And all of the links are given in the description.
Assistir ao vídeo original no YouTube
Assuntos: Agentes de IA, Assistente pessoal. Publicado no YouTube em 2026-10-18.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.