Principles overview

AIBOS rests on 29 principles, arranged across three heights. Each height answers a different question, and each one exists to serve the height above it — nothing at the bottom is there for its own sake.

The three heights

The top height is philosophy: how power gets shared between a human and the AI working for them, and why that division makes sense in the first place. The middle height is the operating model — fifteen principles an owner or a new hire can read and act on without touching a line of code, covered in full in Operating model. The bottom height is the technical build: how the standard actually enforces everything the layers above it promise.

These heights chain together rather than sitting side by side. A P principle exists to enforce an OM principle, which in turn serves something in KF. When a rule in the standard looks arbitrary at first glance, tracing it upward through that chain almost always lands on the philosophy it's protecting.

Every principle has to explain itself

A principle in AIBOS isn't allowed to just state a rule — it has to carry the reasoning behind that rule, written down as part of the principle itself. A rule with no stated reason is one nobody downstream can challenge or improve on.

The standard holds itself to the identical bar. A nightly check scans every principle and every contract in the standard looking for one missing its reasoning, and anything it finds counts as a blocking problem rather than a note for later cleanup. The discipline runs both ways: the standard doesn't only ask you to justify your choices. It justifies its own the same way.

The flagship principles

The receding human (KF1). A human starts fully in the loop on everything, and steps back one process at a time only once that process proves it deserves less supervision — never all at once, and never for free. What's handed over can be pulled back, too, if a handover turns out to have happened too early. The standard's own image for this is a harbor pilot: someone who steers a ship through the hardest stretch of water, then steps off once the crew and the course are both proven safe.

The owner designs, the agent builds (KF2). While something is being designed and built, the owner's intent is the authority in the room — the owner carries real operating experience the system exists to encode. The agent's job is to build it and deliver it with enforcement that actually runs, rather than just reads well on paper. Real risk gets raised along the way, not quietly waved off with "you probably don't need this yet."

Company as Code (OM1). The company itself runs the way software does: a change gets proposed and tested somewhere isolated before it merges deliberately into how the business actually operates — never overwritten by whichever file someone happened to touch last. The standard sums up what this replaces in one line: the end of pitch_v2_final_FINAL.

One rights model (OM7). What anyone — human or agent — is allowed to do comes from the same formula: who they are, crossed with which part of the company they're acting in. The standard states this plainly: the difference between a human and an agent here is one of degree, not of kind. It's what keeps permissions legible as a company grows from one person to a hundred.

Nothing is lost (OM8). Whatever gets built stays findable. Retiring something means archiving it with a marker and a pointer to where it went, never a quiet deletion. As the standard puts it, this durability guarantee is what makes people willing to let go of an idea and explore widely, because losing the work is never one of the risks on the table.

Reasoning is first-class (OM9). Every real choice ships with its reasoning attached, not just its conclusion. A choice with no stated reason is one nobody can later challenge — so the reasoning travels with the decision permanently, as part of the record, rather than something someone has to go ask about.

Scale is the target (OM11). The design goal sits well above one founder directing a couple of agents: twenty to a hundred people, each directing ten or more agents. Working solo is a legitimate starting point, not the ceiling the system was built for — the coordination problems that only show up at real scale are the reason the standard's structure exists at all.

Behavior changes by mechanism (P11). A tidy document doesn't change what an agent actually does under pressure — a hard gate does. This is the principle that keeps the standard honest with itself: a rule only counts as real once there's a script or a check enforcing it, and words alone are never treated as enough.