The unit of work on this site is the ephemeral agent: spun up with a task, a slice of context, and a set of tools; shut down when the task ends. What keeps the output trustworthy is the harness around them. Two auditors are standing infrastructure, wired in as hooks so they run on every edit; the rest are spun up per job.
The practice ladder's harness section describes the pattern: curated context, decomposed work, permissions in code, verification loops. This site runs on exactly that. Below: each one, its role, and what it audits.
Standing agents
The builder
Writes · edits · deploysThe general-purpose coding agent (Devin) that drafts pages, runs research, edits prose, and deploys the site. The agent writing this sentence.
Role- Turns human direction into pages, edits, and deployments
- Verifies its own work: link resolution, citation integrity, served-page checks after every deploy
- Never merges bulk rewrites of the owner's pre-existing documents without asking
- Both audit hooks fire on every edit it makes; no exceptions for the builder
- Human approval for direction, architecture decisions, and publishing
Prose Auditor
Prose quality · hookDetects the statistical fingerprints of unedited AI prose and forces a context-aware rewrite before text ships.
What it audits- All Markdown and HTML prose in the package (
*.md,site/*.html; skips code blocks and itself) - Two severity classes: fix-on-sight (em-dashes, spaced en-dashes) and flag-for-judgment (~20 grammar and phrase tells, e.g.
moreover,delve,it's worth noting,not just X but Y, intensifiers) - Detection is deterministic (a script); rewrites are done by the agent, in context, never by regex
Structural Auditor
Organization · navigation · hookKeeps the site's structure and navigational flow coherent as pages multiply.
What it audits- Nav completeness: every page lists every page, same order; the current page marked exactly once
- Nav hierarchy: the global pages bar must precede chapter navigation, and sticky bars must wrap it, so no local nav floats alone on scroll (regression-tested)
- Link integrity: every internal link and #anchor resolves; no orphan pages
- Page structure: title, meta description, exactly one h1, footer back-link, heading hierarchy
- Judgment pass (per skill): purpose separation, next-step affordances, label honesty, placement of new content
Reference steward
Link rot · quarterlyGuards the citation apparatus against link rot. References are written to survive dead URLs (author, title, publisher, year on each); this agent makes sure they rarely have to.
What it audits- Every external link on every page, classified: ok, redirected, bot-blocked, unreachable, dead
- Wayback Machine: a recent snapshot exists for each source; missing ones are submitted for archiving
- Local copies of PDF sources (regulatory bulletins, survey reports) kept in a gitignored research folder
Evidence Auditor
Fact integrity · on demand + quarterlyVerifies that every factual claim on the site matches its cited source, and that no number stands without evidence. The audit that protects the site's credibility.
What it audits- Every fact-bearing block on every page (numbers, dates, counts, multipliers), extracted deterministically into a worklist
- Cited claims: fetches the source and compares figure, date, and framing; verdicts recorded in a running report
- Uncited claims: sourced, softened, or removed; vendor-reported numbers must keep their "as reported" framing
- Verdict classes: supported / partially supported / unsupported (fixed immediately) / unreachable (cross-checked against secondary coverage)
Ephemeral agents: the research fleet
The AI roadmap was researched by five parallel research agents, each spun up for one question and shut down when it reported. Their outputs were cross-checked, and every figure that made the page carries a numbered citation in the roadmap's reference list.
| Research agent | Its question | What it fed |
|---|---|---|
| Maturity frameworks | What staged AI-adoption models exist, and where is the industry actually at? | The industry-state table; the four-phase shape |
| Use cases and ROI | Which specialty-insurance AI use cases have measured results? | Phase 2–3 use cases; the economics section |
| Competitive landscape | What have named competitors and vendors actually shipped? | The competitor table; build-vs-buy guidance |
| Regulatory | What do regulators expect, by when, in which states? | The governance gates and the regulatory sequencing section |
| Implementation | Why do most AI programs stall, and what do the 7% do differently? | The failure-mode table; the 1:1 adoption rule; the exit gates |
How the harness maps to this site
| Harness layer | How this site uses it |
|---|---|
| Context engineering | House style, tone, and rewrite rules live in the skill files each agent reads before acting; agents see the rules, not vibes |
| Skills | The Prose Auditor and Structural Auditor are packaged procedures: versioned, testable, invocable on demand (/prose-auditor, /structural-auditor) |
| Hooks | Both audit scripts fire automatically after every write or edit to a prose or site file; they run every time, including at 3 a.m. |
| Verification loops | Citation-integrity checks, link resolution, tag balance, and served-page 200 checks after every deploy |
| Human gate | Aaron sets direction, approves architecture and bulk changes, and decides what ships. Consolidations (like the August 2026 harness-into-practice-ladder merge) are owner decisions; the Structural Auditor enforces the rules after them. Slop is cheap; trust is engineered |