---
title: Agentic stack synthesis
tags:
  - wiki
  - syntheses
sources:
  - "[[Agentic coding craft]]"
  - "[[Personal AI agents]]"
  - "[[Model Context Protocol]]"
  - "[[AGENTS.md vs skills]]"
date: "2026-07-17T06:24:38.562Z"
created: "2026-07-17T06:24:38.562Z"
includeInRss: false
---

# Agentic stack synthesis

Working mental model after first full clipping compile.

## Layers

```
Human intent (SoM/ToDos/Projects)
        ↓
IDE agents (Cursor / Claude Code) + skills
        ↓
MCP / CLI tools (QMD, Raindrop, gws, Defuddle)
        ↓
Persistent agents (OpenClaw-class) — optional, high isolation
        ↓
Compiled Wiki + raw Clippings
```

## Rules of thumb

- Session agents for vault surgery; personal agents for always-on ops — don’t conflate.
- MCP for data/tools; don’t force Gmail into MCP if CLI is reliable.
- Skills for workflows; AGENTS.md for invariants.
- Security tax rises sharply once agents get shell + messaging.