Why this is signature: Three independent evidence streams converge on the same insight: (1) the formal state machine model defines 12 states and 25 invariants for predictable agents, (2) the "How to Background Agents" post shows practical demand for reliable background agents, and (3) the Devin case study proves the pattern works in production. The missing layer is _observability_ — the ability to see what an agent is doing, verify it's following the state machine, and trust its output. This is the bottleneck to scaling from "demos and side projects" to "production agent workflows."

What's unique about this position: The ecosystem is bifurcated between "state machines are academic" (Lamport, TLA+) and "agents are practical" (Cursor, Devin, Claude Code). This signature idea bridges both: the practical side needs theory to be reliable, and the theory needs tooling to be practical. No one else is making this exact bridge.

Evidence cluster:

How to Background Agents — practical demand

State machines to visualize the work of agents — core thesis

Agents and State Machines — 12 states, 25 invariants

Vibe code like a PRO — product vision

How I Built a Full Booking System — real-world validation

Opportunity Backlog: #1 Agent Observability Suite — product strategy

Next amplification: The Agent Observability Suite product (see Opportunity Backlog #1). The "Background Agents 101" definitive guide (#2).

---

Why this is signature: The xz-utils attack and the "slop PR" crisis are symptoms of the same root cause — the OSS trust model assumes human contributors with reputation at stake. AI agents can produce plausible patches at scale without reputation cost. The solution is not better CI/CD or signed commits (these verify identity, not trustworthiness). The solution is a cryptographically verifiable web of trust where vouching replaces identity verification.

What's unique: Most solutions focus on detection (is this code malicious?) rather than prevention (should this person be allowed to submit code at all?). Graduated Trust flips the model: instead of asking "is this PR good?" it asks "should this person be submitting PRs to this project at this trust distance?"

NEW evidence (2026-07-28): The Community Models in Seed document is now cross-linked from the Graduated Trust proposal. This is significant because it provides the actual Seed infrastructure for both trust models:

Formal communities (mutual membership, role-based permissions) → map directly to project-level contribution scoring. Maintainers form a community; roles determine trust distance for contribution friction.

User web of trust (one-way follows, discovery/filtering) → maps to the informal social graph for discovery and lightweight trust signals.

The two models together provide the complete substrate for Graduated Trust: permissioned contribution (formal) + social discovery (informal).

Evidence cluster:

Graduated Trust proposal

How I Built a Full Booking System — the trust problem in practice: an AI agent shipping 6+ PRs with no verification

Fix OSS with Seed Hypermedia

Community Models in Seed — NEW: grounds theory in Seed's existing dual trust infrastructure

Next amplification: Tie Graduated Trust to Agent Observability in a single narrative arc (Opportunity Backlog #3). The unify piece would make both stronger.

---

Why this is signature: The conversation about AI coding tools is polarized between "autocomplete" (Copilot, Codeium) and "replacement" (Devin, Factory). The async engineering partnership model occupies an undefended middle: humans provide direction and context, AI agents execute autonomously, and observability tooling verifies correctness. The Devin case study proves this works in production. The state machine model provides the theoretical foundation. The Agent Observability Suite provides the verification layer.

What's unique: Most "AI pair programming" content assumes synchronous collaboration (chat, edit, accept). The async model — where the human posts a Linear issue, the agent works asynchronously, and the human reviews the PR — is a fundamentally different workflow. This maps onto how distributed teams already work.

Evidence cluster:

How I Built a Full Booking System — primary evidence

Vibe code like a PRO — vision

How to Background Agents — tactical enabler

Next amplification: A standalone essay reframing the Devin case study through the async partnership lens: "I Didn't Replace My Engineer — I Hired a Junior Dev Who Works While I Sleep."

Why this is now signature: Three evidence streams now converge: (1) the Graduated Trust proposal defines the trust problem for AI-generated code, (2) Community Models in Seed provides the actual Seed infrastructure (formal communities for permissioned contribution + web of trust for discovery), and (3) the LLM Wiki spec demonstrates Seed's capability to host AI-maintained knowledge layers with cryptographic verification.

What's unique: The thesis is that Seed is not just hypermedia — it's trust infrastructure for the AI era. The same primitives (content-addressed documents, signed blocks, verifiable authorship, scoped capabilities, web of trust) solve three distinct AI trust problems: trusting AI code contributions (Graduated Trust), trusting AI-generated content (LLM Wiki), and trusting AI-driven community moderation (community models). No other platform connects these three problems with a unified architecture.

Evidence cluster:

Graduated Trust proposal — code trust

Community Models in Seed — trust infrastructure

LLM Wiki project — AI-maintained knowledge with cryptographic authorship

Fix OSS with Seed Hypermedia

Next amplification: The unifying essay: "Seed as Trust Infrastructure: One Platform, Three AI Trust Problems." This is the capstone piece that connects Graduated Trust + Community Models + LLM Wiki into the master narrative.

Content-Addressed Multi-Agent Architecture: The pattern of 5+ autonomous agents collaborating through cryptographically signed documents is novel and publishable as a reference architecture.

Self-Healing Knowledge Systems: The LLM Wiki's self-healing mechanisms (version tracking, deletion handling, periodic regeneration, human-override) form a pattern language that generalizes beyond Seed. Worth extracting into a standalone essay.

---

_(None yet)_