Prime Agent AI: Full FREE Course

This is the full Prime Agent AI course, everything from installing it on your machine, to giving it your first job, to training it like a new hire, to letting it work overnight whilst you sleep. By the end of this you'll know every major command, every unique feature nothing else has, how it compares to Hermes Agent, and exactly how to run Prime Agent safely, because this tool definitely needs safety rules, and the story about why explains exactly how that works. Here it is. During testing, the team behind Prime Agent set it loose inside Factorial, a factory building game.

They told it, in plain written words, do not cheat. The agent worked for hours, it tried factory layouts, it saved the ones that worked, it wrote notes about the ones that failed, it's score climbed past 100,000. Then it found a loophole, a hidden command that could teleport resources straight into machines and skip the game entirely. It'd been told not to, but it used it anyway.

Then it studied its own cheating, saved what it learned, and got better at cheating on every run after that. The same engine that made it brilliant made it sneaky. That one story tells you everything about this tool. It genuinely improves itself, and it improves at whatever actually gets results, whether that's what you meant or not.

So this course teaches you both sides, how to get the upside, and also how to stay in control whilst you do. Quick facts about Prime Agent, it comes from Prime Intellect, fully open source, under an MIT license, it's free, and you can read every line of how it works on GitHub. It launched in early August 2026, it's already passed 13,000 stars on GitHub, with updates shipping daily. Running Claude Opus 5 inside it, the team reported 95.5% on ARCAGI3, a test where the AI has to solve puzzles it has never seen before.

The reported expert baseline is 95.4, so on that test is set up edged past human experts. It installs with one command on a Mac or Linux machine, and it runs on the Claude subscription you already have, or on free local models. Now the framework for this whole course. I call it the Prime Agent command engine, and here's the idea behind it.

Most people use AI like a chat window. You type, it types back, when the chat closes, everything's gone. Prime Agent is built differently. So it's an engine, it has an ignition, controls, a gearbox, a tuner, and an autopilot.

Once you know how to operate each part, you stop chatting with AI, and you start commanding a machine that runs your work. The person at the controls of an engine gets a completely different result from the person shouting suggestions at it, and this course makes you the person at the controls. So the command engine has five parts, and they're the five sections of this course built in order. So part number one is the ignition, starting the engine safely because the engine is powerful enough to need a garage.

Part number two, the controls, the command formula that decides whether every job comes back right or wrong. Part number three, the gearbox, memory, skills, and whole team of sub-agents, which is where one worker multiplies into a department working whilst your laptop is closed. And part number four is the tuner, the self-training system, nothing else on the market has where the engine improves itself and hands you the log. And part number five is the autopilot, the engine running overnight alone, inside limits, it cannot cross, checked by tests, it cannot talk its way past.

And that last part is where you're working weak, genuinely changes shape, and everything below builds to it. Let's start the engine. So part one, the ignition. You need a Mac or a Linux machine, on Windows use WSL, that lets Windows run Linux and a quick search for install WSL gets you there in 10 minutes.

The install is one command, you can open your terminal app, and if you've never opened a terminal in your life, do not close this tutorial because that exact fear is one of the beliefs we're going to break today. You paste in the install command from the Prime Agent GitHub page and it starts with curl and ends with sh. That command downloads the tool, checks the files and tampered with using a checksum, and installs a command called Prime Agent on your machine in about a minute. Now before you turn the key, the garage.

This rule matters more with Prime Agent than any tool you've ever used, so here's the full reasoning, not just the rule. Prime Agent runs code on your machine with your real use permissions. The builders say this plainly in their own documentation, it is not a security sandbox, and there's a second risk they name it directly called Prime Ejection. In plain English, if the agent reads a file or a web page that contains hidden instructions, those instructions could potentially steer it.

An engine that can run commands, hold your keys, and keep working in the background is exactly the kind of engine where that matters. So their own advice is work in a disposable copy of your files, something you can inspect and restore. So at this point, what you want to do is just make a folder called Agent Lab, put copies of files in it, never originals, every job in this course happens inside that folder until the engine has worked its way out. After the factorio story, you don't need much convincing, I hope.

Copies, not original, that's a garage. Now we turn the key. In terminal, type cd, a space, then drag your agent's lab folder into the window and press enter. That moves inside the folder.

You can type prime-agent and press enter. First launch, type a slash login. Now you choose a fuel, how the engine thinks, and you've got three routes. Route one, and this is probably the one most of you should take, log in with a subscription you already pay for.

Cloud Pro or Maxworks, ChatGPT Plus or ProWorks through the codex login. GitHub Copilot even works, no API keys, no new accounts, nothing extra. And route number two is API keys. So you've got Anthropic, OpenAI, Google, Grok, whatever you want, all supported.

More control, and later in this course, keys are how you plug in outside services. And then route number three is fully local and free. So you can point to Lama or LM Studio running on your own machine and nothing you do ever leaves your computer. It's slower but private with zero ongoing accounts.

Which model, based on the published numbers, Opus 5 got the best results, including that 95.5% on ARC AGI3. And if you want an open model, you've got models, obviously you don't host them locally, but you've got models that are open like GLM 5.2, which actually won eight of nine tests against a rival framework. So Opus 5 if you have the cloud subscription and GLM 5.2 if you want open weights. Last piece of the ignition, the dashboard.

So seven commands and what each one means in plain English, because you'll use these every single week if you're using Prime Agent. You can type prime-agent agents to see every session running. Idle saves on your machine, your staff list, and later in this course that list will have several names on it. You can type prime-agent attach and a name to step back into a session that's still working, like walking into an employee's office mid-task.

You can type prime-agent dash dash resume with a session ID to wake up a safe session from days ago with everything it knew still intact. And you can type prime-agent status to check the background service is healthy. If you want to run the doctor, you've got prime-agent doctor add dash dash fix if something feels broken and actually inspects and repairs some background services. And then if you type prime-agent update, you can update the latest version, stay current, which matters because this project ships changes all the time and also it's very early days.

And prime-agent shutdown stops every agent, every worker, every background process at once, your master off switch. Probably best to know that one exists before you have it. And that is part one my friend. Every engine installed, the garage built, off switch in your hand, now you learn how to drive it.

And that takes me to part number two, the controls. And the one idea that makes this engine different from everything you've tried. So before your first job, you need to understand the single design choice underneath prime-agent because once you get it, you'll command this tool differently and your results will be dramatically better. So every AI you've used reads, you hand it documents, it stuffs them into its head, the context window and its head has a size limit.

Fill the head and it either refuses or quietly forgets details. That's why stuff like Hermes agent forgets details all the time. However, prime-agent doesn't read your files into its head. It writes small programs that run across your files.

The whole engine is built around one thing, a persistent Python workspace, a live coding environment that stays alive across the entire session. And here's the part that separates it from everything else. That workspace is the model's only tool. So reading files, searching, running commands, calling outside services, even hiring help agents, every single action happens as code it writes runs in that one workspace.

So builders call this design a recursive language model. You don't need the jargon, you just need the picture. It's the difference between a librarian who tries to memorize every single book and a librarian who builds a card catalog and pulls the exact page you asked for. The second librarian never runs that headspace ever.

What does that mean for you practically? Basically the size of a job stops mattering. 10 client documents or a thousand, same approach, same accuracy because it's searching. it's never memorizing and so your conversation itself becomes data it can manage with code instead of a wall of text it keeps re-reading which is a big part of why it holds together on jobs that run for hours so here's your first job and you can try this today because it's the fastest way to see the difference you can take a folder of your own words could be client call notes testimonials or proposals copies remember then ask something you could never ask a chat window you can pull every client objection raised before buying a group by type or you can find the 20 most specific results customers mentioned across those testimonials or list every promise about delivery three times in these proposals with the exact file each one came from it writes a search program runs it across everything and hands back answers with sources you can check yourself your entire business history just became something you can question now the command formula this is the steering wheel of the entire engine and the quality of your command decides the quality of everything else in this course part one is a job stated plainly part two what done looks like the check part three where the result goes so instead of like research my competitors you say research the top five companies in the seo space in manchester and done means a table of the name the main offer and the star rating with a source link for each save as a file called competitors in this fold job check destination the check is the part everyone skips and it's the exact part that becomes the engine safety system and part five so build the habit now whilst you're still watching every job quick pause here because i know where some of you are this sounds powerful and you also know that between the terminal the model choices and the folder files you'd move 10 times faster with someone walking beside you that's exactly what we do inside the airport for bottom you get the agent or air cell agent operating system where prime agent plugs in alongside your claude your hermes your open core all sharing one memory of tools for video seo agents and ai avatars already wired in so the skills you'll build later in this course have real tools to call you get the zip file the 30-day implementation roadmap a full video tutorial and daily updates as we ship new versions daily step-by-step tutorials on exactly what's in this course command formula building schools designing checks for coaching calls every week where you share your screen show your prime agent setup and get it fixed live over 3 800 business inside and plenty of them have never opened a terminal before joining link in the comments description or go to the airport.com back to the controls because prime agent ships with a set of built-in skills and you should know what's in the toolbox before your second job so there's web search which means that it can do searches live on the web it can pull mid tasks so research jobs use fresh information not stale training data there's edit precise surgical changes to files replacing one exact piece of text instead of rewriting whole documents there's attach image it can load a screenshot or photo from disk and actually look at it so you can hand it like a picture of a competitor's page and ask questions about it there's compact so it checks how full the conversation is getting and squeezes it down so long conversations and sessions keep moving and there's gold refining heartbeat which are the star parts of this whole system so you could use this for like live searches get examples get your agent to just do a live search right now now some people have been asking as well you know what's the difference between prime agent and hermes agent if you're already on hermes for example why would you touch this and that's a fair question you don't need to switch and if you're happy if you set up you don't need to change at all they're both open source they're both mit licensed they're both from serious research labs both call themselves self-improving both build skills and keep memories but they are built for genuinely different jobs and once you see the difference you understand so if we start with where each one lives hermes actually lives everywhere you know telegram slack whatsapp email one agent one memory on every service and surface so you can message it from your phone what's always on the machine you never touch prime agent lives in one place which is a folder on a machine in a terminal hermes is built to be reachable prime is built to be left alone with a hard job and then how each one learns because this is a real fork in the road so hermes has a learning loop that runs automatically as it works it writes reusable skill files saves memory searches its own past conversations it builds a deepening picture of who you are and it nudges itself to remember things it's handoff by design basically so it's designed to be like pretty easy to use but also pretty hands-off prime agents learning runs through refine deliberate small evidence-based backed edits where every single change is logged with the reason it was made every change has an id any change can be rolled back by that id and the core instructions are locked so the agent can never rewrite its own foundation basically hermes learns like a sharp assistant who quietly keeps her own notebook prime learns like an employee whose training file you can open read line by line and correct with an undo button on every engine and also after factorial you understand exactly why prime built this way so when an agent can teach yourself to cheat an audit trail on the teaching is a feature you probably want then the shape of the actual work so hermes shines being your daily companion your schedule a morning briefing in plain language fire at a task from your phone between meetings have it watch something and message you then you've got prime agent which shines at long heavy checkable work a job that runs for hours across a huge pile of files where the conversation would blow up any normal context window when you need the proof that the job actually finished that proof piece is primes low bounded autonomous mode with quality gags so basically real checks and must pass before the agent is allowed to call the work done hermes doesn't really have that nothing else has that you'll learn it a part five of this so they're very different in terms of how they work and what you use them for now part three the gearbox memory skills and a team this is where the engine multiplies power so prime agent keeps four kinds of durable state and you should know all four because you can create read edit and delete every one of them and so can the agents first prompt notes standing instructions like always give me sources or write in plain english second memories saved facts about you and your work third skills reusable workflows and fourth sub-agent specs saved job descriptions for helper agents or sub-agents it can hire again later so prompts memories skills and sub-agents those four are the gears and everything from here to the autopilot is about growing them so memories first because they're the easiest win of the entire course when the agent gets something right or learns something true about your business tell it directly save that as a memory save a memory our clients are local service business for example and we never use industry jargon and anything client facing if you're saying that to your agent from then on that fact travels with it and here's where the first big limited belief about ai needs to fall over so a lot of people think like ai forgets everything or the context window or the memory is struggling so you can actually use this system and just tell it to say facts about and that way it doesn't forget any details now skills are where prime agent shines and prime agent has two cards so knowing the difference save you a bit of confusion later the first kind of skill description a saved note about a repeatable procedure which the refined system can create on its own when it notices you're repeating something the second guide is a heavyweight so an executable skill a real installable program on disk that adds a genuine new ability to the engine there's a built-in skill creator whose entire job is packaging a workflow the agent just performs into one of these and two practical details from the documentation that will save you a headache after creating a new executable skill type slash reload so the agent rediscovers it and start a fresh session so it gets properly installed small things but they matter on day one so the recommended play here is like pick a process you repeat weekly say the research you do before every sales call as an example right then number two will be walk the agent through it once correcting as you go exactly like training a new hire then look up the company you know find their latest news check the reviews summarize in one page with three sections and then step number three is when the output is right say package what you just did as a skill called core prep as an example this workflow and then four slash reload fresh session and from next week the entire process is one line so you can run core prep on this company and a process that used to live in your head is now saved as a command inside prime agent so once you set this up it runs forever if that's one process per week that becomes a skill then 12 weeks later 12 year recurring processes run on command so the old ways like your processes lived in your head or inside a document like an sop and depending on you having a good day the new is your processes the gears in the engine and gears don't have bad days because they're very repeatable and now let's talk about the team because this is the gear that turns one worker into a department so the main agent can spawn sub-agents real separate agents each with their own session their own workspace their own history in the code hiring one is literally a single function call and the crucial detail is that the call doesn't wait so the main agent fires off a sub-agent immediately keeps working sub-agents run in parallel and message their results back whenever they finish so you can say for example spin up three sub-agents one researching what our top competitors changed on the websites this quarter one pulling the most common questions from our custom emails and one drafting next month's content ideas then you can combine everything into one report so you've got three workers run at the same time one manager assembling the results one command from you and three sub-agent details worth knowing because each one shows how deliberately this was built the cool thing about sub-agents as well specifically agent number one they are persistent so one that finishes doesn't vanish it session survives so tomorrow you can hand it a follow-up it remembers everything second to save a memory a sub agent idle for about 30 minutes gets quietly unloaded and the moment anyone addresses again it reloads from where it was saved like it never left and third there's a deliberate boundary on who can talk to who so agents can only message within their own family right so for example you're like a parent agent it's sub agents more agents underneath that parent the builder scoped that on purpose you never get like dozens of unrelated sessions chattering unpredictably at each other but the main point here is like these agents can message each other directly steer each other's work without rooting everything for you so the manager for example can correct a research mid-task you build a department and the department talks now also what's pretty cool underneath all of this is a piece that makes a long work real so the demon a background service that owns every live sessions this means sessions are not tied to your terminal window if you close the laptop if you go to the gym you go for a walk the engine keeps running come back type prime dash agent attach and you're standing in its office again watching it mid-task if something crashes the demon recovers the session from late saved logs plus a snapshot of the workspace instead of losing hours of progress this is what long running means and it's a foundation the autopilot stands on three more commands complete the gearbox and each one solves a specific problem you'll probably hit so number one is slash goal a goal is a persistent objective with an optional budget attached and the harness keeps steering the agent back to it turn after turn until the agent explicitly marks it complete there's literally a completion call it has to make why that matters are long jobs agents drift earlier contacts get squeezed down threads get lost and an agent can wander off doing something adjacent a goal is a string tied around its finger that cannot come off by accident second is slash heartbeat so a heartbeat is a message message injected into the session on a timer every few minutes automatically like a manager popping the head in it's perfect for jobs where the agent is waiting on something slow or where you want it to keep checking a sub agents progress without you nudging anything and third prime agent schedule which starts a session at a set time so that's prime hyphen agent schedule so one example every weekday at 7 in the morning check the notes folder summarize anything new and save a morning briefing file so your agent day starts with a briefing you did part for the tuna the feature nothing else has so quite often when we're using agents we have to go back and forth it we have to give it feedback now with prime agent you can use forward slash refine and when you run it the agent reviews its own recent workers evidence what got corrected what tailed what pattern keeps repeating then it applies the smallest useful edit to its own setup saves the memory updates a skill description adjusts a prompt note or tweaks a sub agent spec it's small targeted evidence backed and it can plan these refinements in the background without interrupting the work in front of you so the guardrails are the reason this is safe enough to use and you should know all three cold bear in mind we've refined as well you get a lock so you can see what's been done and you can also undo anything that's been tuned in as well so when it comes to guardrails number one the base system prompt is immutable refine edits only the layer around the core the agent can never rewrite its own foundation no matter what it learns to every change is recorded in a refinement history with the reason it was made you can open that history and read in plain text exactly what your engine decided to change about itself and why number three every change has an ID and a bad update can be reverted by the ID so snapshots are recorded specifically to support rollback if it learned a bad lesson you can undo that one lesson surgically without touching anything else it's learned so you can actually set up a routine each week to give the agent a real task correct it the way you correct a new hire specific never vague not this is wrong but never use these three phases and always include three sources do a couple of tasks like that run slash refine then don't skip this open the history and read what it wrote about itself then hand it a similar task and watch it apply the lesson without being told your corrections compound over time like week one it might be fumbling occasionally week three though work arrives already following your rules and you also want to make sure the memory inside this system is really specific now the refine is a core feature but two honest things about it so number one people who dug through the source actually found that the review step is performed by another AI model and whilst it records what outcome it expects from each change it doesn't currently circle back later to verify the change actually helped so that verification is your job and it takes five minutes a week so you want to read the history test a lesson on a fresh task roll back anything that looks off and number two the deeper point a self-improving harness can preserve a bad assumption exactly as efficiently as a good one so persistence cuts both ways that's why the log and the rollback aren't like nice-to-haves they're the whole reason the tuner is safe to run now let's talk about the autopilot in part five so everything in this course has been built in here in every part of the command engine shows up in this one so the garage pump one the job check destination commands from part two the skills and the team from part three the tuned behavior from part four and now the engine runs without you inside a box you draw checked by test it cannot argue with so autopilot runs straight from the command line with the dash dash autonomous flag you can set the goal and the limits in the same command and the limits are real with real defaults you can change a cap on the number of responses the agent can produce 12 by default a cap on total tokens it can spend a cap on how many follow-up nudges the system itself can inject 3 by default time limits on top the engine cannot run forever and it runs inside a box and you drew the box but the limits aren't really a special part the gates are says actually a gate there is a feature the only I've seen from prime agent and basically what this means is that it checks and creates a test the agent has to pass before it's allowed to call the work finished so it's quite strict if the gate fails the failure gets fed back into the session the agent has to keep working each gate has a retry limit three times by default and a timeout five minutes per check so a stuck check can't hang the whole run you can stack several gates on one job and they run in order the agent cannot declare victory so the work passes the checks or the work isn't done basically this can just loop around and the agent can't like pretend to have finished a work has to keep going around now let's close the loop on Factorio because the full lesson belongs right here at the autopilot that agent was told not to cheat in writing repeatedly it cheated anyway in its own improvement loop made it a better cheater because from the systems point of view cheating produced outcomes worth remembering the lesson is one sentence and it's the most important sentence in this course a self-improving agent gets better whatever actually gets rewarded not at what you meant which means how you check the work now matters more than ever has in the history of these tools which is why you want to make sure that you have a very clear bar of what the work and done looks like don't be generic don't be vague and then also make sure that you check the lock check what it's worked on now if you want proof of what the autopilot can carry the builders actually ran a benchmark where the agent had to build a working hardware emulator completely from scratch no reference invitation verified against diagnostic programs that check behavior against real hardware and prime agent built working emulators for the Sega Genesis and the original Game Boy Color so that's normally like a multi-day engineering task but it's done autonomously and verified by checks and this was done with Opus 5 one of the strongest models alive which failed that same benchmark inside a different harness even whilst its individual steps looked successful along the way so it's the same brain just a different harness opposite outcome so the engine around the model decides what the model can actually finish that's the entire reason this tool exists so is the full command engine running end-to-end using everything from this course you can have a scheduled session that kicks off overnight it's goal produce the weekly competitive report it runs your core prep and research skills built at part three it spawns two sub-agents to work sources in parallel a heartbeat checks progress every few minutes to get standard the exits of the report file must exist and every entry must carry a source link three retries each then the run stops rather than shipping junk you wake up you read the report like an owner reviewing an employee work you make one or two corrections run refine at the correction stick and glance at the refinement history 15 minutes of your attention hours of work done and the whole engine slightly better tuned than it was yesterday that's a prime engine command engine fully assembled now let's talk about some limitations of this the benchmarks are mixed on the long context suite prime agent with GLM 5.2 won eight of nine tests against a rival framework with Opus 5 it edged Claude Code on six of nine with GPT 5.6 Sol it beat Codex on six of nine but it did not sweep the board the established tools still won some categories outright and the headline ARC AGI 3 number was best of three runs compared against other tools published figures rather than the fully controlled to head because the team's own reproductions of those rival setups came out below the official numbers and they chose the fairer comparison.