synthesis-autopilot

Synthesis Autopilot — Autonomous Execution Mode

The Problem

Users who delegate whole tasks to an agent end up retyping the same paragraph of standing instructions every time: complete all the phases, don't check in constantly, use my decision framework, keep a plan file, don't lose state when the context compacts, build the real solution rather than a workaround. Retyped instructions have three failure modes. They drift — each retelling drops a clause, and the dropped clause (usually verification, or the decision protocol) silently doesn't happen. They decay — a long autonomous run outlives its own instructions when the context window compacts, and the agent reverts to conservative defaults mid-task. And they don't compose — the instruction block names disciplines that live in separate skills, and a paraphrase of a skill is weaker than the skill.

This skill encodes the delegation contract once. One explicit phrase from the user engages the mode; the mode sequences the skills that already exist rather than restating them.

What This Is — and Is Not

A mode. Activation changes the agent's check-in cadence and self-management for the current delegated task: fewer questions, batched questions, a plan file that survives compaction, verification before "done." It does not change the quality bar — the anti-shortcut and integrity disciplines apply to all work, supervised or not.

A thin composition layer. Every discipline this mode invokes is defined in its own skill, listed in depends_on. This file sequences them and states the protocol that is unique to autonomous runs (trigger discipline, the plan file, batched decisions, alerts). When this file and a dependency appear to disagree, the dependency is authoritative for its own domain.

Not an authority expansion. Autonomy governs how the agent sequences work and how often it interrupts — never what it is permitted to do. See "Standing Gates Survive Autonomy" below.

Activation — Trigger Discipline