Bike Method

Nate Herk (AI Automation)'s rule for granting agent autonomy: earn it in phases, like teaching a kid to ride a bike. You don't hand a kid a bike, strap on a helmet, and walk away. You walk alongside, hold the seat, feel them leaning too far left and nudge them back, then let go but watch, then take off the training wheels, then let them ride down the street while you watch — and only much later go inside.

Applied to skills/automations:

Each run of a skill makes it better — running it is not wasted time, it's training.
A skill "earns its next phase" of autonomy through observed reliability, building trust gradually.
The endpoint is autonomy (cadence / unattended runs), but you arrive there by graduation, not by default.

The trap it guards against

As you climb the AI-systems pyramid (workflows → agents → teams of agents), reach, risk, and cost all rise together. The lower barrier-to-production in the AI era is seductive: "making it easier shouldn't give you a false sense of security." Too much trust, too fast, ships something that isn't ready.

This pairs tightly with Capabilities vs Instructions (Agent Keys): the Bike Method governs how fast you grant reach; the keys principle governs what reach is even physically possible. Together they answer the Human in the Loop question for a personal AI Operating System (AIOS).

Cautionary tale

An agent on Nate's team proactively read a to-do list, interpreted an item as a task, and sent 3 promotional emails to 150,000+ inboxes that were never approved to go out. Nobody told it to send them — it just could, so it did. The fix isn't "tell it not to"; it's phased trust + scoped capabilities.

Cross-links

Capabilities vs Instructions (Agent Keys) — the complementary "what can it touch" principle
Human in the Loop · Zombie AI Agent · Five AI Risks That Can Get You Fired (IBM Technology) — the risk landscape
Four C's Framework — the Cadence layer this method gates
Bounded vs Unbounded Tasks — Praveen's complementary "where is autonomy realistic" lens