Autonomous Development Skills Suite: Four Composable Skills for AI Agents

What this is. An implementation of the Autonomous Agent Principles framework as four composable operational skills for autonomous agents.

The four skills.

1. Intent - surfaces the agent's interpretation of the ask before any work begins. Runs first; Improve and Probe delegate to it when installed.
2. Improve - an examine, decide, change, verify, record loop that can either make one high-leverage change or conclude that nothing should change.
3. Probe - paired tests that distinguish reasoning about the specific situation from pattern-matching against a generic template.
4. Trail - appends a structured, multi-resolution evidence entry to trail/log.md after every session.

What else is included. An append-only trail format plus Python tools for recording entries and mechanically verifying integrity.

Why it matters. The suite is designed to address two recurring problems with coding agents: they do not know when to stop, and they can produce plausible generic fixes that are not truly grounded in the target codebase.

Evidence status. This release is the first with full convergence evidence: three independent silence runs from distinct model families plus a cross-layer coherence check.

Read in this order.

1. README.md - overview, scope, and evidence status.
2. intent/SKILL.md - run first.
3. improve/SKILL.md - the default improvement loop.
4. probe/SKILL.md - the reasoning probe.
5. trail/log.md - the recorded evidence.

Canonical source: GitHub repository. Companion theory: Autonomous Agent Principles on Zenodo.