How it works

You set a heading and hold it, correcting course as conditions change, instead of steering toward a fixed point and hoping. The system follows the same idea: durable state instead of one-shot prompts, completion that is checked instead of assumed, and routines that tell you when you drift.

Most agent setups keep code and data in one place. HEADING OS splits them on purpose.

The engine reaches your data through one seam, and seven mechanical layers keep the two apart. Security lists them.

The command post is your main session. Everything live happens there and nowhere else: mail, calendar and messages, the background services, every push of your repositories, and the review and merge of finished work.

Agents each work in their own copy of the engine, on their own branch. Several run at once. A guard stops any of them from sending, from starting or stopping a service, or from publishing anything. When an agent finishes, its work is reviewed and merged in the command post, and its copy is removed.

The rule that decides which side an action belongs to is one question: does it touch something live? A running process, an outside party, or a shared index other work reads. If yes, it belongs to the command post.

An executive never does one thing at a time, and the assistant should not either. In HEADING OS each larger task gets its own yard: a separate workspace where one agent takes that job from start to finish. Three yards can be at work while you sit through a two-hour meeting.

• You keep working. Mail, the calendar and your decisions stay in the command post, and work in progress never touches them.
• Nothing half-finished goes live. A yard can research, write and build, but it cannot send, publish or change anything running.
• One mistake stays in one yard. If a task goes wrong, you drop that yard and lose nothing else.
• You review once, at the end. Each yard hands back its result with a note on what was done and what was not, and it joins your system only when you say yes.

In practice this is a small team that works through the afternoon on the jobs you handed out, while you do the part only you can do.

Most dropped work starts as one sentence: "don't forget to chase Northwind after the board call". In HEADING OS a sentence like that becomes a task the moment it is said, whether you said it to the assistant or an agent noticed it in the middle of its own job. Before writing it down, the assistant checks that the same task is not already there.

Open any card and you see its whole life. Here is task 31 by the end of the same day:

• It lands in your inbox, not on your plate. A new task waits there until you approve it. Deciding what is real stays with you, and only you drop a task.
• Every task says who has it. An agent that takes one writes its name on it, and when it finishes it writes what was done and what was not.
• Order is kept for you. "After the board call" is recorded as a dependency, so the task comes forward only when the thing before it is finished.
• Done means done. A task closes when the work it names is finished and accepted, never because a plan said so.
• The record stays. Tasks and their full history live on your machine, beside the rest of your data, and a new session starts by showing you what is waiting for your decision.

So the question "did anyone pick that up?" always has an answer you can read, for your own tasks and for every agent's.

The engine ships 82 skills: drafting mail in your voice, meeting preparation, a file-based CRM, research and intelligence briefs, decision analysis, branded documents and decks, a weekly review. You do not need to remember their names. A router matches what you say to the right skill; measured on 2026-10-15, it chose correctly in 660 of 677 test cases.

A model forgets when the conversation ends. HEADING OS does not rely on it to remember.

• Decisions, threads, contacts and notes are files in your data repository.
• A semantic index runs on your machine, so a question like "what did we decide about pricing" is answered from the record, with the file it came from.
• Before a session is cleared or compacted, its state is written down and handed to the next one: the task, the constraints, what was approved and what was not.
• An optional memory server can sit beside the index for recall across everything you have worked on.

Mail sync, backups, index refreshes and health checks run as scheduled services that survive a reboot. A watchdog notices when one stops. A radar lists the routine work you have let slip, such as a backup or a weekly review, and tells you, not anyone else.

• Windows 11 with WSL2, macOS 13 or later, or Linux.
• An Anthropic account: a Claude subscription, or an API key.
• Comfort with a terminal, and about 60 to 90 minutes for a first install.
• Some skills need more: a mail account, local models for private work, or keys for research services.

If that list is the reason to stop, installation and support are coming.