Agentic OS by AgentTaker
The operating system for teams of AI agents.
One person runs a whole team of AI agents in parallel, safely, from a laptop and a phone. Planners compete, a judge picks, a worker builds, a fresh checker tests it, and a guard blocks money actions in every mode.
- 20 offices
- 219 office runs in 19 days
- 252 Brain skills
- Claude, Gemini, NVIDIA NIM or local Ollama
Counted from the code and run folders on 25 and 26 Oct 2026.
See it working
The real dashboard, at work.
Or open it on YouTube: youtu.be/KMXb3KED6Lk
- Jack, the main agent, takes your ask and routes it to the right office or pane.
- Offices plan, judge, build and check a job, with evidence.
- The guard, Freeze and tap cards keep money, sign-ins and stop buttons with you.
- The Brain keeps lessons and quizzes the offices on them.
For a company of 5 to 50 people
One AI office for the one process that keeps you up at night.
In many small firms the owner or one engineer does the paperwork at night. Slow paperwork loses money: a missed tender, a late quote, a patient who does not come back. You need one more person but cannot hire. We set up an AI office on that one process.
-
Contractors
Tenders on time
An RFP arrives with a short deadline.
- Compliance checklist
- Technical proposal and method statement
- Company profile, kept current
Your engineer checks and submits.
-
Clinics
Patients who come back
Follow-ups slip when the front desk is busy.
- Reminder and follow-up drafts
- Replies to reviews
- 12 posts a month and a monthly report
Your staff approve before anything goes out.
-
Trading and import
Quotes the same day
Supplier prices change faster than the catalogue.
- Supplier comparison in a day
- Quote drafts from your price list
- Weekly price report, catalogue upkeep, follow-ups
Your sales lead checks and sends.
The rule in every pilot: AI drafts, a person in your company checks and sends. You bring your own Claude plan or API key and pay the provider directly; we never resell model use.
Offices
Every job goes through an office, not a single chat.
An office is a team of agents for one kind of work: research, code, media, security, jobs and more. There are 20 on the laptop today, each with its own leaderboard and lessons. Jobs run in parallel; a swarm of 2 to 6 agents can split one job, each in its own git worktree.
- Research
- Web
- Media
- YouTube
- Security
- Product
- Data
- Jobs
- Blog
- Marketing
-
Three planners compete
Each one plans in a forced thinking style, and one of them is a random wild card, so the obvious plan does not win by default.
-
A judge picks one plan
The judge reads all three plans and picks the one to build.
-
A worker builds it
Parallel jobs lease the files they edit. A second job that edits a held file gets a warning naming who holds it, so nobody silently overwrites anybody.
-
A fresh checker tests it
Pass, fail or unsure, with evidence. A fail gets up to three fix rounds, an unsure gets one send-back, and a failed run can be rolled back.
Safety
What stops it spending money or going wild?
Under the offices sits a guard on every tool call. The money block and Freeze hold in every permission mode. The lease and trifecta checks warn and list; they do not block.
-
Money block
Payment actions are blocked in every permission mode. The guard fails closed: if a check errors or passes its 7-second limit, the action does not run.
-
Freeze
One button stops every agent: no tool runs while it is on. Only you can unfreeze, in the browser.
-
Phone tap cards
When only you may press something, like a sign-in or an “Authorize” button, a card with a picture of that one window goes to the dashboard and your phone. AOS never presses it.
-
Per-file leases Not found*
Parallel jobs lease the files they edit. A second job that edits a held file is told who holds it.
-
Lethal-trifecta check Not found*
One command lists every live agent that holds private data, text from strangers and a way out, with its three legs.
-
Number check in code Not found*
An AI summary with a figure that is not in the given facts is dropped or flagged by code, not by another model’s opinion.
-
Audit log
The guard keeps a log of its decisions, and one command checks it:
aos audit verify. Full mode allows and logs; anything on the NEVER list still goes to a card. -
Tool trust list
Connected MCP tool servers are on a trust list. If one changes its tools, you get an alarm.
-
Keys and files stay safe
No keys in files: the Vault hands a key over for 30 seconds, only after you allow it. Cleanup moves files to Trash; nothing is deleted.
Try the button: it is a picture of the real one. *Not found = not found among the 19 agent products in our two scans of 10 Oct 2026 or on the competitor pages re-opened on 23 Oct 2026. It is not a claim that nobody in the world has it.
Brain and recall quiz
A memory that is tested, not just saved.
- Lessons keep their reason and source. Each rule carries why it exists, where it came from and its votes. Rules you type win over learned ones.
- New skills are tested first. A skill is tried with and without, then waits for your OK.
- Nothing is deleted. Old lessons stay, so you can see what changed and why.
- Quizzed after training. An office answers questions on what it learned. A miss goes on a “Remember first” list in every later job until a later quiz passes it. Not found*
What the quiz found
The offices saved lessons but did not use them. Quizzed on its own lessons, the media office got 2 of 12 right. After a short “Remember first” list went into every job:
The YouTube office went from 7 to 11. Saving memory is not the same as using it, so AOS now tests what each office has learned.
Take the quiz yourself
Three questions on this page, scored the way AOS scores an office.
Voice, phone and Telegram
Talk to it. Check it from your phone.
-
Live talk
Speak to Jack, the main agent, with a wake word or by holding Ctrl+Space. The Live tab is never recorded.
- English
- Urdu
- Hindi
-
Telegram
Send a job or a voice note to a Telegram bot and get the result back. Voice notes are written out with Whisper.
-
Phone app
The dashboard on your phone over a private Tailscale network, never a public link. Tap cards land here.
-
Steer while it works
Steer a running job, answer an agent’s question card, or ask about any card on the board.
Voice runs on Gemini Live, with Whisper, Kokoro and Vosk for offline speech.
Any model
Not tied to one model vendor.
Each pane is a real Claude Code session with its own model, effort and mode. One switch moves the main agent to another provider.
- Claude CodeHaiku, Sonnet or Opus. The default.
- Gemini CLIGoogle’s models.
- Claude APIPay per use with your own key.
- NVIDIA NIMHosted open models.
- Local OllamaThe last step: runs on your own machine.
Plan-pace guard. When 5-hour plan use reaches 80%, an Opus office job waits and other Opus jobs step down to Sonnet. Jobs stopped by the weekly limit carry on after the reset.
Everything in it
Every feature, in plain words.
Run the work
- Jack, the main agent, routes every ask
- Panes: real Claude Code sessions, each with its own model, effort and mode
- Live board of every agent and job
- 20 offices with leaderboards and lessons
- Swarms of 2 to 6 agents, each in its own git worktree
- Workflows and Automations, with an eval gate
- Tasks timeline with rewind, and save points
- Six layouts and a Ctrl+K command bar
- A fresh session takes over when context passes 60%
Stay safe
- Guard on every tool call that fails closed
- Money block in every mode, 7-second limit
- Freeze: only you undo it
- Phone tap cards for sign-ins; AOS never presses them
- Per-file leases and the lethal-trifecta check
- Audit log with
aos audit verify - MCP trust list with an alarm on tool changes
- Click guard and send receipts: no repeat send within 3 minutes
- Vault for keys; cleanup goes to Trash only
Brain
- Rules keep why, source and votes
- Typed rules win over learned ones
- Skills tested with and without; they wait for your OK
- Recall quiz and “Remember first” list
- Number check in code
- 3D galaxy view of everything it knows
- Nothing is deleted
Talk and watch
- Live talk in English, Urdu or Hindi
- Wake word and Ctrl+Space push-to-talk
- Steer a running job
- Show screen, answer cards and explainers
- Ask about any card
- Telegram bot and phone app
- Voice notes written out by Whisper
Keeps itself running
- Self-upgrade every day; asks before it installs
- Self-heal watchdog: up to 3 fixes per thing per hour; never deletes, pushes or sends
- Tokens page and the plan-pace guard
- A PC worker over SSH
- MCP servers for tools
Your data stays yours
- Local-first: it runs on your own computer
- Phone access over a private network, never a public link
- No keys in files
- Any model, including a local one
Proof, honestly
A working product, used every day.
- lines of JavaScript
- 93,506
- offices
- 20
- office runs on the laptop, 8–26 Oct 2026
- 219
- Brain skills, 46 of them with test cases
- 252
Counted from the code and run folders on 25 and 26 Oct 2026.
Where we are: the founder uses it heavily every day. No paying customers and no outside users yet. Next: three pilot companies, each with one AI office on a real process.
Pricing
Start small. Pay more only when it works.
Free call
SAR 0
30 minutes. We ask four questions to see if an AI office fits.
AI office check
SAR 1,500
about USD 400
We measure the real numbers on one process before you pay for a pilot. Comes off the pilot price.
Pilot
SAR 5,000
about USD 1,333
30 days against a success test we write down together first.
Monthly
SAR 3,500 / month
about USD 933
The office keeps running, and we look after it.
First 3 companies: SAR 2,500 a month.
Founder
Built by a founder who uses it every day.
Muhammad Ahsan
Founder · AI engineer · Riyadh
- 2023 · Degree in Artificial Intelligence
- 2023 → · Building AI products online
- May 2026 · Started Agentic OS, full-time since
“I build everything with AI coding agents. Running many at once kept failing the same ways: they overwrote each other’s files, stopped at sign-ins, could leak data or spend money, and forgot what they had learned.”
“I started Agentic OS in May 2026 to fix that for my own work, and I use it every day. I write all of the code with AI coding agents.”
Pilot waitlist
Join the pilot waitlist.
We are looking for three pilot companies, each with one AI office on one real process. Leave your details and we will write to you about a pilot place.
- We read every sign-up ourselves.
- If your process fits, we offer a free 30-minute call.
- No newsletter. We only write to you about the pilot.
Contact
Tell us about one slow process.
A free 30-minute call. Investors are welcome to write too.
ah@ahsanawan.com