Status: 🔥 Active — material exists across multiple short posts; needs synthesis

Thesis: State machines are not just for UI. They are the right abstraction for making LLM-driven agent behavior deterministic, observable, and composable. The same XState patterns that power Seed's document machine should power agent workflows.

Evidence:

State machines to visualize the work of agents — direct argument

Notes about how the document machine should work — real state machine implementation in production

Modeling resizable panels using State Machines — UI application

State machines are great, but hard to establish — the adoption challenge (which is itself an opportunity)

How to Background Agents — connects the state machine model to Cursor's background agent pattern

Why Horacio owns this: He has a production state machine (Seed's document machine), a public argument for state machines in agent contexts, and a unique vantage point building both the collaboration platform AND using AI agents extensively.

Investment required: A flagship article or talk synthesizing these threads. Currently all pieces are scattered across short posts.

---

Status: 🔥 Emerging — Devin post is the strongest signal yet

Thesis: The best AI coding partner isn't the one that writes the most code. It's the one that works like a skilled engineer you direct asynchronously — scoping, breaking down work, checking its own results, and communicating clearly. This is different from "vibe coding" and different from "AI replaces developers."

Evidence:

How I Built a Full Booking System as a Side Project — primary case study

Vibe code like a PRO — envisions the Linear→Agent→PR pipeline

How to Background Agents — practical background agent usage

Why Horacio owns this: He's a product-minded engineer building a product (Seed) while using AI as a partner — not a tool. His angle is unique: he treats AI like a junior-to-mid engineer who needs context and direction, not like a magic code generator. This is an antidote to both "AI hype" and "AI skepticism."

Investment required: Turn the Devin post into a series. A follow-up on "prompting for refactors" (the Vitest→Bun story deserves its own article). A third on "architecting with AI" (the parent-child account management story).

---

Status: 🟢 Active — implementation in progress

Thesis: The document editing lifecycle is a state machine with 10+ states (loaded, editing, publishing, rebasing, etc.). Publishing this as an open-source case study shows how XState v5 powers real collaborative editing.

Evidence:

Project Plan — New Publish Mental Model Pt.1 — 465 lines of XState v5

Opportunity 14: Document Machine OSS Package — extraction plan exists

Case Study: Document Editor — external educational resource already exists

Why Horacio owns this: He designed and implemented it. Few people have built production state machines at this scale for collaborative editing.

---

Graduated Trust for open-source in the AI era — interesting but needs more real-world validation

Codesigning Electron apps — practical but not differentiated

Type pairing for readability — interesting niche but not signature material