cortexOS: A 24/7 Daemon That Runs Coding Agents While You Sleep

The video opens with the creator talking to camera in a bright apartment, a bold caption pinned at the top: "YOUR DASHBOARD IS NOT AN AGENTIC OS." His argument is that the wave of slick Claude Code dashboards people are shipping right now does one thing -- it kicks off a single task and shows your usage limits. That is a control panel, not an operating system.

He then cuts to his own build, cortexOS. The screen recordings show a real web app with a left-hand nav (Overview, Agents, Tasks, Activity, Context, Comms, Approvals, Workflows, Strategy, Analytics, Members, Experiments, Skills) and a separate Telegram view where each agent appears as a named contact. The pitch is that the difference between a dashboard and an agentic OS is what runs when the browser tab is closed.

The Test: Can Your Agents Run Without You?

His opening line is a checklist. If your Claude Code and Codex agents "don't work for you overnight, you can't send them tasks from your phone, they can't talk back and forth to each other, you do not have an agentic OS." He grants that the popular setups look good -- one-click skill that deploys Claude Code, completes a task, shows usage limits on a clean UI. His point is that those are bounded, one-shot interactions. Nothing persists. Nothing coordinates.

He says he ran coding agents at that level for a while and "just was not happy with how autonomous they could behave." So he built a layer on top that manages any coding agent you already use in an interactive session, but keeps it running around the clock.

What cortexOS Actually Shows

The Overview frame is the clearest piece of evidence. It reads "AGENTS ONLINE 9/9, All systems go" and "TASKS TODAY 64." Below that is an ACTION REQUIRED block listing "3 pending approvals" and "10 blocked tasks," a left column titled Agent Fleet, and a right column titled LIVE ACTIVITY streaming events like telegram_received, telegram_sent, and task_completed. This is not a single agent. It is a roster.

The agents have names. Across the dashboard and the Telegram view you can read Boris, Data, Donna, Goggins, Nick, Paul, Sentinel, Skoollo, Tallybot, and a context-designer. The Comms page shows direct channels like "admin to paul," "admin to codex-test," "admin to boris," each with timestamps and unread counts, plus a Meeting Room and "Active Channels 12." One message in the thread reads "admin: send cloudflare link to dashboard," followed by an agent walking through setting up a tunnel. The agents are reading instructions and replying with steps.

The Pieces That Make It an "OS"

He lists the components that run behind the dashboard:

A 24/7 daemon. He describes "a 24/7 daemon that actually manages agents, that keeps them with 24/7 uptime." This persistent process on the machine is what holds state when no one is watching.

A shared task board. "The agents of course have a shared task Kanban board where they can push tasks through to completed throughout the day autonomously." The Overview's 64 tasks and 10 blocked tasks are the visible output of that board.

Schedulable workflows. The Workflows page is titled "Manage scheduled crons across all agents" and shows "TOTAL CRONS 49." He confirms this on camera: "I have currently 49 scheduled workflows." The list includes named jobs like heartbeat (every 4 hours), github-trending-feed, competitive-intel, and research-digest. These are crons living inside the daemon that fire tasks to the agents on a schedule.

Self-improvement loops. The Experiments page is labeled Autoresearch and shows "ACTIVE CYCLES 24, RUNNING 12, COMPLETED 58, AVERAGE 88%." He says the agents use "Karpathy's auto research method to actually experiment and improve upon their deliverables." The page lists research cycles and proposed experiments, for example one proposing to break pending tasks into smaller single-focus chunks to test throughput. The agents are running structured experiments on their own work.

Phone-First Control via Telegram

The last claim is mobility. He shows a Telegram client where every agent is a chat thread -- Paul ("3 approvals still pending, check dashboard"), codex ("Done, I found the native Codex goal support"), Nick ("Posted: Instagram, YouTube"), Boris ("Path B landed"), Tallybot ("CRASH: tallybot died unexpectedly"), and BotFather confirming a new bot. He says you message agents back and forth from Telegram, spin up new agents, and manage all of them remotely: "you never even have to come to your computer." The pitch is that a true agentic OS is reachable from your phone, not tied to a desktop tab.

The Distribution Model

He closes by saying the entire system is hosted for install in his Skool community (the dashboard even has a "Skool" nav item), the link is in his bio, and community members are contributing to the repo and improving it over time. So this is both a product demo and a recruiting pitch for a paid community.

Key Takeaways

• His definition of an agentic OS is operational, not visual: overnight execution, phone control, agent-to-agent messaging, and many agents running at once. A dashboard that fires one task and shows usage limits fails all four.
• The load-bearing component is a persistent 24/7 daemon on the machine that holds state and keeps agents online (the Overview reads 9/9 agents online), which is what a browser-only dashboard cannot do.
• cortexOS coordinates a named fleet (Boris, Data, Donna, Nick, Paul, Sentinel, and others) through a shared Kanban board, with 64 tasks shown for the day and 10 flagged as blocked.
• Scheduling is concrete: 49 cron-style workflows like heartbeat, github-trending-feed, and competitive-intel fire tasks to agents on a timer.
• The agents run self-improvement loops he attributes to Karpathy's auto-research method, with the Experiments page showing 58 completed cycles at an 88% average and proposed experiments on their own throughput.
• Control is phone-first through Telegram, where each agent is a chat thread and you can message, approve, or spin up new agents without touching the computer.