Two Paths to Done

Once a piece of work is scoped, there are two legitimate ways to actually carry it out, and neither one replaces the other. The choice is a judgment call about how much a human needs to look at each individual step versus how much an automated check against a confirmed goal is enough judgment for that particular stretch of work.

The phased path

A plan is researched and written up first, then implementation proceeds one phase at a time, stopping for an explicit human review and approval at every phase boundary. This is the right choice when the value of a human eyeballing each step outweighs the time cost of stopping to ask that often — genuinely high-stakes or unfamiliar work, or anything where the plan itself is likely to need correcting mid-flight.

The autonomous loop path

A vision and a concrete, checkable definition of done are confirmed once, up front. After that, the work iterates — implement, verify against an automated quality bar, self-check against the confirmed goal — without stopping for approval at every individual step, until the goal is met, a pause checkpoint is reached on schedule, or a genuine blocker comes up. The trade is explicit: less per-step human review, in exchange for running much longer unattended stretches. It is not a lower-quality version of the phased path — it's a different allocation of where the human judgment goes.

Neither path skips the gate

Whichever path is chosen, any action with real, hard-to-reverse consequences still stops for its own named approval gate — the autonomous path removes the per-phase review, not the gate that sits at the actual boundary of doing something irreversible.