Agentic Engineering Patterns
A Claude Code plugin for building software products with AI agents — from raw idea to shipped MVP.
Why This Exists
Traditional software development bottlenecks on human coding time. Process design optimizes "how to make people write code faster."
When agents can execute dozens of tasks in parallel, that bottleneck vanishes. A new one takes its place:
Agent execution capacity is near-infinite. Specification quality is not.
Vague specs don't slow down a human — they ask a colleague and adjust. Vague specs paralyze agents — they guess, diverge, and produce incompatible code across parallel sessions. The cost of ambiguity scales with parallelism.
This inverts the entire design logic:
Traditional: plan roughly → adjust as you go → ship
(optimizes for human coding speed)
Agentic: invest heavily in spec precision → parallel execution → ship
(optimizes for agent execution quality)
Every skill in this plugin serves that logic. The time you spend in /envision and /map pays back exponentially when agents build in parallel without asking questions.
The Mental Model
The workflow separates thinking from doing:
┌─────────────────────────────────────────────────────────────────┐
│ │
│ CONTROL PLANE (human decides what to build) │
│ │
│ You + AI collaborate on high-leverage decisions: │
│ goals, decomposition, architecture, priorities, feedback │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ /envision │───►│ /map │───►│ /reflect │──┐ │
│ │ │ │ │ │ │ │ │
│ │ what to │ │ how to │ │ what we │ │ │
│ │ build │ │ break it │ │ learned │ │ │
│ │ │ │ down │ │ │ │ │
│ └──────────┘ └──────────┘ └──────────┘ │ │
│ ▲ │ │ │ │
│ └────────────────┼──────────────┘ │ │
│ │ feedback loop │ │
│ ▼ │ │
│ ┌────────────┐ │ │
│ │ /dispatch │ picks stories │ │
│ │ │ from the map, │ │
│ │ what to │ creates OpenSpec │ │
│ │ work on │ changes │ │
│ │ next │ │ │
│ └─────┬──────┘ │ │
│ │ │ │
└────────────────────────┼─────────────────────────┼──────────────┘
│ │
story specs │ status + cost flow up │
flow down │ │
▼ │
┌──────────────────────────────────────────────────┼──────────────┐
│ │ │
│ EXECUTION PLANE (agents build it) │ │
│ │
│ Agents receive precise specs, work in isolation, │
│ produce PRs. They don't decide what to build. │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌─────────┐ │
│ │ /design │───►│ /launch │───►│ /build │───►│ /wrap │ │
│ │ │ │ │ │ │ │ │ │
│ │ refine │ │ spawn │ │ implement│ │ archive │ │
│ │ the spec │ │ agent │ │ + test │ │ + update│ │
│ │ │ │ │ │ + PR │ │ status │ │
│ └──────────┘ └──────────┘ └──────────┘ └─────────┘ │
│ │
│ (repeat per story — multiple stories run in parallel) │
│ │
└─────────────────────────────────────────────────────────────────┘
Agents don't talk to each other. They communicate through structured artifacts — context documents, story specs, interface contracts, signal files. The harness coordinates everything. This is a production system design, not a chatroom-style agent swarm.
The Plugins
Each plugin implements one layer of the mental model.
1. Product Context — the persistent map
memorysaver