Architecture Overview
agent-config is a content + governance layer, not an agent runtime. The agent loop, the model dispatcher, and tool orchestration all stay with the host tool; agent-config supplies the artifacts and the discipline those tools read.
The four artifact types (+ personas)
Section titled “The four artifact types (+ personas)”| Artifact | Count | Role |
|---|---|---|
| Skills | 278 | Intent-matched capabilities and executable procedures |
| Rules | 105 | Always-active behaviour constraints, auto-loaded per conversation |
| Commands | 190 | Slash-command orchestrators that compose skills + gates |
| Guidelines | 101 | Reference conventions, read on demand |
| Personas | 29 (+5 advisors) | Review lenses dispatched during code/design review |
Skills carry YAML frontmatter and follow the Agent Skills open standard. Rules split into an always-on kernel and trigger-activated tiers — see Kernel & Router.
Authoring-time vs runtime
Section titled “Authoring-time vs runtime”The scope line is authoring-time (in scope) vs runtime (out of scope): code that runs during prompt construction or on an explicit user command — assembling context for the agent to review — belongs; code that executes autonomously during agent operation does not.
What it deliberately is not
Section titled “What it deliberately is not”- No background daemon — the host runs the loop.
- No separate state database.
- No self-rewriting memory — every learned change is human-reviewed.
- No auto-build pipeline.
Capability without a process to babysit — and every such claim is machine-checked.