IntrinsicOS

IntrinsicOS is the vault the company runs out of. Not a metaphor: every project's state, every open task, every decision and its rationale, and every prompt any agent is ever sent lives as a plain Markdown file in one folder. A person reads and writes it in an ordinary notes app, from a laptop or a phone. A crew of named agents reads and writes it on a schedule.

What it does, mechanically

Nine times a day, six days a week, a scheduled tick runs one model session over the whole folder. It consumes the inbox — an unorganized brain-dump file — and routes each item to the right project. It checks each project's stated position against its code repository's history and corrects drift. It rewrites the human's to-do list to hold only the things that genuinely need a human. Around that one judgment pass sits a deterministic chassis that does the boring, checkable things — syncing check-offs, taking snapshots, committing, pushing — without a model in the loop.

Separately, an executor picks up tasks that are marked as delegable and runs each one as its own session in a git worktree, on a branch, ending in a pull request. Whether that PR lands by itself or waits for a human depends on a single line in the project's file:

autonomy: gated

locked means draft only. gated means build to a boundary — a deploy, a device test, a client send — then stop and ask. open means run to completion. A project with no line never executes, and the rule is enforced in the code that spawns sessions, not in the prompt.

Why it's an instrument

Every one of those sessions writes a run report. Every review gate records what it found. Every autonomy refusal, every held task and the reason, every decision that had to go to a human — all of it is logged in the same plain files. That makes the vault a continuous record of what agents actually do when the work is real and the stakes are someone's actual company, which is the dataset the agent-operations research area is built on.

The review gate

Automatic landing requires a passed independent review of the final diff. The reviewer starts without the author’s conversation and examines the changes against the task requirements. Blocking or unavailable review holds automatic landing; unfinished or unreviewed work can still be preserved in a held pull request. Publishing that branch is not evidence that the work passed review or is ready to release.

Measuring progress toward autonomy

Our goal is a company that can carry work from an idea through delivery with far less day-to-day involvement from its founder, while meeting or exceeding his standard of quality. IntrinsicOS lets us study that question through real client and product work.

The home page's operating measures show a dated snapshot derived from the vault's records:

• Recorded session completion compares completed sessions with attempted sessions only where both counts were recorded and valid. Missing or inconsistent counts are excluded, with incomplete coverage disclosed. A session can finish at a review or delivery boundary without the work being accepted by its stakeholder.
• Work aimed at clients and products counts sessions with a written report, treating every project other than the vault itself as outward work. It measures dispatch share, not time spent or value delivered.
• Choices that passed without veto counts choices recorded as proceeding after the founder’s veto window, as a share of all dated decision-log entries. It does not establish how much he shaped a choice beforehand, and an entry without that record does not establish his involvement either.

Documentation drift is tracked separately by comparing the Arbiter's reported stale claims with the claims it checked. It is a documentation measure, not an autonomy score.

The next stage of this research follows whole pieces of work through delivery: whether the result is accepted, what needs correcting, how long it takes, and how much of the founder’s time it requires. That will help us understand where agents can take greater responsibility and where better judgment or tooling is still needed. The current figures describe individual parts of the operation, each with its own source and observation window; they are not yet an overall autonomy rate.

What is not released

The code, for now, and everything in the folder — it is one company's operating state. What we publish is the design, the measurements, and the failures.