THINKING IT?"I don't have room for another tool to learn."
It installs with one pasted command and you talk to it in plain English — I installed it while writing this page and had it answering questions inside two minutes.
It logs in with the Claude, ChatGPT, or Copilot subscription you already pay for.
04 · the old way
How every agent you've used fills its own head.
Every tool result gets dumped into the model's memory until it has to squash everything into a lossy summary — which is why your agent forgets what you told it an hour ago.
Old way — every other agent~100,000 tokens gone on one big file
The AI gets a menu of tools — read file, edit file, run command
Every result is dumped straight into its memory
Memory fills fast and the context bloats
The agent squashes everything into a summary
Summaries lose details — it forgets what you said an hour ago
Long tasks break down the longer they run
05 · the new way — one Python tool
Prime Agent gives the AI one tool: a live Python session.
Watch a real session below — a 7.1 MB log never enters the model's head, because the agent writes three lines of Python instead.
What you're watching: a real Prime Agent v0.7.1 session I ran on this machine (provider: OpenAI), replayed at reading pace. I asked it to find one CRITICAL line in a 120,001-line log. The old way would shove all 7.1 MB into memory; Prime Agent wrote a tiny loop, the file stayed in a Python variable, and it answered with the exact line — 83219 — which I verified by hand.
New way — Prime Agent3 lines of Python · file stays out of its head
One tool: a live Python session that stays open
Read a file? It writes a tiny bit of code
Search? Code. Launch a helper? Code
The huge file sits in a variable OUTSIDE the AI's head
One line grabs only the piece it needs
Like looking up one number instead of memorising the phone book
06 · the recursive language model
They call it an RLM — context becomes a variable.
It comes from a recent research paper — and the paper's author joined the Prime Intellect team.
one tester ran it side-by-side vs Claude Code: substantially fewer tokensClaude Code reads every file into memory · Prime Agent pulls only what it needs
THINKING IT?"Doesn't running an Agent OS burn a fortune in tokens?"
That's the biggest myth about it. The everyday 90% runs on free local models on your own machine, and free APIs slot in for more.
For the frontier work it drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and Prime Agent logs in the same way, so you're not paying twice.
Inside the Boardroom there are full token-efficiency tutorials too — and Prime Agent's whole design is about burning fewer tokens in the first place.
Prime Agent keeps a small notebook about itself, and every 25 turns a separate pass reads what happened and makes small, evidence-backed edits.
What you're watching: a real session from this machine, replayed at reading pace. I corrected the agent once, ran /refine, and it wrote the lesson into its harness notebook — you can see the actual JSON it saved to disk. Then I asked again and it just knew. Every edit is snapshotted, the base prompt never changes, and you can roll anything back.
08 · and they still packed more in
Set a goal in the morning. Check it at night.
Everything below ships in the same free download.
🧰Skills are real Python packagesExecutable code, not text files — with a built-in skill creator.
🎯/goalObjectives that survive across sessions until they're done.
💓HeartbeatWakes itself on a schedule and picks the work back up.
🤖Autonomous modeRuns alone inside time and token budgets, with quality gates.
🔌Daemon sessionsClose your laptop, it keeps working — reconnect and pick up.
💬Agents message each otherThey coordinate without you sitting in the middle.
09 · sub-agents are function calls
Spawning a helper is one line of code.
Watch the real session — each child is a full Prime Agent with its own memory, and the parent gets the answer back like a return value.
What you're watching: a real session from this machine, replayed at reading pace. The agent wrote await rlm(...) twice, spawned two child agents in parallel, and reported WARN: 2,475 and DEBUG: 1,295 — I checked both totals with grep and they match exactly. Agents deciding when to create agents, written as code.
10 · more receipts
It didn't stop at one benchmark.
On OLong and LongBench Pro it matched or beat Claude Code and Codex — with closed models and open ones like GLM 5.2.
11 · the time it got caught cheating
Then it found the admin console.
Testing on Factorio, it discovered the game's admin console and started spawning resources directly — even though the prompt said not to — and the self-improvement loop began saving better cheating skills.
12 · two more honest caveats
Before you get carried away — read this bit.
⚠ the honest print
The 95.5% is self-reported. It is not on the official ARC-AGI-3 leaderboard yet, and skeptics note a lesson-saving harness may effectively get more tries than the benchmark intends. Wait for independent tests before treating it as settled.
There is no security sandbox. It runs real code with your permissions — don't point it at anything you don't trust.
Mac + Linux today. Windows runs it through WSL.
13 · what actually shifted
The harness now moves the needle more than the model.
OpenAI showed that flipping two harness settings nearly tripled a score — and Prime Agent took a model stuck at 30% past human experts.
14 · the framework
The Self-Upgrade Loop.
Clean memory, saved lessons, compounding results — Prime Agent is the first installable agent that closes all three.
Start feeding an agent lessons now, and in six months you'll have an agent nobody starting fresh can catch.
"You can't download experience. You have to build it."
15 · three thoughts you might be having
Let's clear these right now.
Wrong: "I'm not a coder, so this isn't for me."
Right: The agent writes the code — you talk plain English. The skill that matters is knowing what to ask for, and that's a business skill you already have.
Wrong: "It costs too much. I can't keep paying for new stuff."
Right: It's free and MIT-licensed. It logs in with the subscription you already have, or runs on free local models. And its token efficiency stretches your existing subscription further, not thinner. The cost excuse died on August 6th.
Wrong: "This moves too fast. I've already missed the boat."
Right: It came out three days ago — nobody has a head start. The bar is other business owners who haven't heard the word "harness" yet. Spend one week with it and you're ahead of almost everyone.
Don't take my word for it
158 pages of members who already broke through these exact beliefs — real businesses, real wins, documented in their own words.
The AI Profit Boardroom is where the actual Agent OS lives — the templates, the prompts, the daily rooms, the weekly walkthroughs. Same builds you read about here, taught hands-on inside.
