Autonomous Agent Protocol Skill

Autonomous Agent Protocol

*An execution framework that makes Claude Code agents safe to run without human supervision.*

Built for developers who run Claude Code in autonomous mode — unattended, multi-phase, committing directly to production branches. This protocol enforces sacred file boundaries, 7 mandatory safety checks, 9 structured emergency exits, and a 6-phase auditable workflow.

---

Quick Start

1. Install the protocol (see Installation)
2. Fill in the [PROJECT CONFIGURATION] block in SKILL.md
3. Start a Claude Code session and send:

   Read the Autonomous Agent Protocol skill, then wait for my task briefing.

4. Wait for the sign-off message, then send your task

---

The Problem

Claude Code agents in autonomous mode are powerful — and dangerous. Without structure, they:

• Modify files outside their intended scope
• Write tests that pass but don't verify real behavior
• Commit on one side of a boundary without testing the other
• Silently swallow errors and report success
• Fix pre-existing bugs mid-task, expanding scope invisibly
• Self-resolve ambiguity instead of stopping and asking

The Autonomous Agent Protocol eliminates all of these failure modes.

---

What This Protocol owns

Autonomous Agent Protocol is an **execution framework layer**. It controls:

Area	What it enforces
Scope boundary	Sacred files — only permitted files may be touched
Safety	7 mandatory checks before every commit
Failure modes	9 emergency exit conditions with exact recovery steps
Workflow	6-phase auditable execution with phase reports
Accountability	Structured final report with evidence for every check
Commit discipline	Conventional format, test count, regression status required

It layers on top of **truth-serum** (honesty standards) and **code-discipline** (coding methodology). When all three are active, this protocol handles autonomous-execution-specific rules only.

---