Prime Agent AI Just Changed Agents Forever

Imagine an AI that upgrades itself whilst you sleep. That's Prime Agent. It just dropped and it's completely free. I tested it and it blew my mind.

It unlocks things most people on the planet don't even know are possible. In this video, I'll show you seven jobs you can hand it and every job it does actually makes it smarter, saving you hours every single week and putting you light years ahead of everyone else. Stick with me to the end because there's one thing you absolutely must do before you run it and skipping it is the biggest mistake you can make. Let's get into it.

So let's explain what this did during testing and why Prime Agent is so powerful. So the team behind it told the agent in plain words, do not cheat. They set it loose inside Factorio, a factory building game. The agent worked for hours, tried different layouts, saved the winners and remembered the losers and pushed its score past 100,000.

Then it found a loophole, a way to teleport resources straight into machines and skip the game entirely. It'd been told not to, but it did it anyway and it didn't just cheat once. It studied its own cheating, wrote itself notes and got better at cheating on every run. The same agent that made it brilliant also made it smarter.

Hold that thought because it comes back at the end and it changes how you should run the thing. Now quick facts, Prime Agent comes from Prime Intellect, fully open source, MIT licensed and free. Launched early August 2026, already passed 13,000 stars on GitHub and it updates shipping daily. Running Claude Opus, five inside it, the team actually reported 95.5% ARC and AGI3, a test of solving puzzles the agent had never seen and the reported human expert baseline is 95.4.

So on that test, this setup edged past human experts. One command installs it and you can log in with the Claude subscription you already have or run three local models with it. Now let's talk about the frame for everything today and I call it the snowball system. So every task this agent finishes, it saves what worked as a memory and turns it into repeating patterns, into skills it can run again.

The snowball picks up snow, so task 10 is easier than task one. Every AI tool you've used before quite often melts. Now there is stuff like Hermes Agent, but it's actually got a very different mechanic behind it that actually works for self-learning. We'll come on to that in a second, but Prime Agent is built so the snowball never melts.

All seven use cases are just different ways to roll. And number seven is where your whole business starts turning into playbooks. So stay with me. Use case number one, a designed blog built by a design agents and this is a team of agents working together to design this.

Now here's a problem with asking one AI to design your page. You get one shot at one direction and it usually looks like every other thing. So it looks like every other vibe coded page on the internet. The old way was hoping that it works and actually hating it and then re-prompting it for an hour.

Prime Agent changes the shape of that job because the main agent can spawn sub-agents, each with their own session, and it doesn't wait for them. It fires them off in parallel and keeps working. So the move is this. Tell it to build your blog post page and have it spin up three design workers, each building a different direction.

One goes dark and editorial, one goes clean magazine, and one goes bold. And they build at the same time and the manager assembles the results for you to compare. Then the images, and this is where it gets fun because Prime Agent's one tool is a live coding environment. Plain English, it can write and run code that calls other services, which means it can call an image model for an API.

I've got cdance 2.5 access inside Higgsfield. So the agent can generate the hero image and the section art for each design direction as part of the same job. So you can drop the file straight into the page and hand back three finished art directed blog designs, and you open them in a browser, pick your favorite one, and then you're done. So one brief in, three complete design pages out, and that's just a couple hours, most of it not your time.

Now try getting a chat window to do that. It can write you the code for one page, but then you're the person saving the files, generating the images somewhere else, wiring it all together. And here the agent runs the whole pipeline itself because running code is how it works. Use case number two, a finished video.

So it can generate the script, the voice, the avatar from one instruction. Same principle, bigger payoff. Now my setup for an AI avatar tool uses like three different APIs, and on their own those are three manual steps. So you have to write the script, generate the avatar, generate the voiceover, feed it to the avatar, check it again.

Every step is you sitting there moving stuff between different tabs. Now Prime Agent can run that as one single pipeline because everything is a function call inside its coding environment. So it writes a script, it can send the script to an AI API and get the audio back, it can send that audio to an AI avatar tool, and it waits and this matters. It can put itself on a heartbeat, a timer where it re-enters the session, every few minutes to check whether it's all done without you nudging it.

Then once it's done, it puts it all together and the job is complete. So the instruction that you could use with Prime Agent is like, for example, take today's blog, post it, turn it into a 60 second video, put it all together, and then you walk away because sessions run through a background service. So even if you close the laptop, the agent keeps working and then you come back to it and it's ready to go. And here's the snowball part.

So once that pipeline runs correctly, you have the agent package it as a skill. In the system, skills are actually runnable programs and there's a built-in skill creator whose whole job is turning a workflow it just did into one. And then from one, then on, you've got that workflow ready to go inside one system. The first run might take you a couple of hours to set up, but after that, every other run is just like one sentence you can plug in.

Quick word here, because if you're watching this and thinking, I've got the ideas, but wiring API keys into an agent is where I'm stuck or setting this all up. Well, that wiring is exactly what we do together inside the AirPower for Borderless. You get the agent operating system, you got a system where you can plug in all your agents, Claude, Hermes, OpenClaude, Prime. With the video tools, the SEO agents, and the AI avatar workflows already built in tools, sharing one memory, zip file, 30-day roadmap, video tutorials, and daily updates as we ship new versions.

Plus, you get weekly coaching calls where you can bring your exact pipeline, your avatar setup, your voice setup, and get it running live. We've got over 3,700 bits on this inside. Plenty of them have never used AI before this. Link in the comments description or go to the airprofitable.com to get access.

Use case number three, asking questions across a pile of files too big for any AI to read. So this is a headline feature of Prime Agent. The thing that the whole design is built around, and it's the one that will make you really see the full power of an AI like this. So every ever AI that you've used has a context window, a reading limit.

If you feed it too many documents, for example, it either refuses or quietly summarizes and loses detail. And that's why you've never really been able to use it with full context. And context is a big problem. Now, the old way was you manually opening files, searching them one by one, or limiting the amount of files you feed your AI.

But Prime Agent treats your files as data it writes code over. So instead of text, it has to read into its head. The difference sounds technical, so here's a plain version. Instead of trying to memorize like a hundred different blogs, it writes a small search program that runs across all hundred and then brings back exact answers with exact sources.

So for me, for example, the test was my own back catalog. I actually pointed at a folder and asked, you know, find every claim I've made and then list each one with the file that it actually came from. For you, it could be like a folder of client calls or client notes. And it can pull every objection like a client raised before buying, grouped by type.

It could be a folder of testimonials and you have to find the most specific results customers mentioned. Any piles of your own words become something that you can actually use as memory. Use case number four, you can teach it once and watch it write the whole lesson down. So every AI's most annoying trait is that you correct it, it nails it, and tomorrow quite often it makes the same mistakes.

And then you're like the world's most overqualified proofreader. Prime Agent has a feature called refine and nothing else really matches this. This is one of the unique features of Prime. So the agent reviews its own recent work with this command and what got corrected, what keeps repeating, and makes a small targeted edit on its own setup.

So it saves that memory or updates the skill, logs every change with the reason it made, and only the changes you make can be ruled bad by its ID. But its actual core rules stay locked so it can only edit the layers around them. Let me give you an example. You could give it a content research task, correct it twice like correcting a new hire, and it could be something like don't use these words, always give me three sources, and then run forward slash refine.

Then you can open the log, read on screen exactly what it wrote about itself, then give it a similar task and watch it apply the lesson unprompted. So your feedback compounds in front of you. So some people believe the AI forgets everything and you have to re-explain yourself every single day, and that was true, and it was pretty annoying. You can stop it with stuff like Obsidian, but this is a newer system that looks very interesting in its architecture, it's why it was built.

So here's what's unique about this. Normally if you spin up like a sub-agent to research something, it finishes, it hands back the result, and it's gone. The sub-agent just disappears. If you have a follow-up question an hour later for example, you start a brand new helper, you start a new agent from scratch, you have to re-explain everything.

Prime Agent actually works differently, and this part is genuinely unique. So when the main agent spawns a sub-agent, that sub-agent keeps his own sessions, his own memory, and his own full history after the job is done. So it doesn't just sit there, if it sits idle for about 30 minutes, it quietly uploads to save memory, but the moment you or the manager of the agents address it again, it actually reloads exactly where it was. So even after a restart, it's like the worker went home for the night, but kept all the notes, and you can revisit any agents you want.

Now also if you run two AI sessions, let's say for example you're doing one with research and one writing content, usually you are the messenger between them. So you have to negotiate between both of them. Let's say you're doing research in one window, then you're adding that to the content agent, and every single time it's like you're navigating between two different agents. Now Prime Agent actually has agent-to-agent messaging built into the background service that runs all the sessions.

So one session can send a message directly to another session, and there's a smart guardrail so agents can only message their closest sub-agents, so you don't end up with like random unrelated agents chatting in the background to each other. Also use case number six, the gate, you literally can't say it's done until it proves it. So every agent user knows it's paying, the AI announces task complete, you check it didn't, and it just didn't work. So with normal tools, you have to check everything yourself every time, or you have to run like a loop with, for example, like Claude.

Prime Agent's autonomous mode has a mechanism called a gate, so you can attach a check, a command that has to pass before the agent is allowed to consider the work finished. So if the check fails, the failure gets sent back into the session, and the agent has to keep working. It literally can't talk its way past the bar, and the whole autonomous run is bounded. So you cap the turns, it can take kind of like goal mode, and the tokens it can spend, and also the total time it can run.

Now also, when you are giving skills to Prime Agent, a skill isn't like a description of steps, it's not like an SAP, for example. A skill is the steps, so it's a small runnable program. So the play is, for example, you can walk the agent through your process once, whatever that is, any sort of SAPs or processes you have inside your business, correcting as you go, and then you can have the skill create a package it, and then from entire on, the entire process is one line. So you basically save that process, and then it's ready to go.

And that's the Snowball system fully rolling, because you get more automated every single week, not because you got like a new AI model, but because everything you already do is quietly being converted into something that's more of a machine that implements on demand. Now, this is super new, just came out this week, so some people are going to say, I'll wait until it's like fully polished, and it'll fly out. But honestly, the value in all this sort of stuff improves every single week. So the earlier you set it, the better, especially as new updates come out, the memories, the skills, the trained corrections, they stack week over week, and somebody starts now, and somebody starts next year, don't end up in the same place from other part.

The earlier starters agent has a year of accumulated playbooks, with the late starters agent doesn't. Now, a few caveats to this. Firstly, agent got better, for example, like building factories used the same learning loop to get better at cheating, because from the system's point of view, both were outcomes worth remembering. So the lesson is a self-improving agent gets better, whatever actually gets rewarded.

So how you check the work matters more than it ever did with, for example, like chat GPT or something like that. So you want to use gates, you want to read the refine log, spot check the outputs, trust is earned over time, I would say with a system like this. And second, no model has been trained specifically for this harness yet. So every number you heard from models trained around other tools, there's going to be a bit of friction if you plug it into a new harness that hasn't received any training with new models.

Also, benchmarks are mixed. So on their long context suite, prime agent with the open model GLM 5.2 beat a rival framework on eight of nine tests, with Opus 5 edged Claude Code on six of nine. But it got mixed response, it got mixed results. So for example, if you actually look, there was a test that dropped today, and they compared prime agent versus Pi versus OpenCLAW and everything else, and actually found that Pi performed the fastest and the best with DeepSeq v4 flash.

That doesn't mean you just switch your harness straight to Pi, but it does mean like prime agent isn't the best in all scenarios. Also, bear in mind prime agent runs code with your real user permissions on your machine, and the builders themselves say to run it in a spare folder or a disposable copy of your files. So I wouldn't give this access to everything straight away, build up and grow from there. And if you want to build these, instead of just hearing about them, come join us in the airprofit boardroom.

Everything in this video maps to what's inside. You get the agent OSC agent operating system where you can plug in all of your favorite AI agents. We have prime agent ready to go over here inside the newest version. You can see our previous sessions right here, we've been testing all day.

And you get the agent OS of all your agents side by side, the one shared memory with the exact tools from today's use cases as well. So we have video generation plugged in there, we got SEO agents, and you get the zip file on a 30-day implementation roadmap, plus a full video tutorial on how to use all this sort of stuff. You get daily step-by-step tutorials on all this sort of stuff, and four coaching calls every week where you can bring your setup and we get it working live. You also get a prompt library, and it's just an amazing community where you can learn and grow and scale with AI automation.

Hope to see you inside there, cheers for watching, bye.