A content operation run by AI agent loops

The problem

A consulting practice needed a body of work deep enough to answer the specific questions multi-location operators ask: one page per operational problem, per audience, per implementation. Written by hand, that is years of work.

Volume alone was never the goal. Generated copy drifts, repeats itself, and invents facts. The real problem was producing at volume while holding a quality bar a human reviewer would sign.

What was built

Agents run in recurring loops, each with a narrow job: draft a page family, rewrite against the house doctrine, add a measured internal link, or audit a defect class. Each loop works from a backlog file and a registry, so its state lives in the repository rather than in a conversation.

Every change ships as a reviewable pull request, stacked when one depends on another. Nothing reaches production without a human merge.

The quality bar is enforced by code. Each defect class found in review became a check that fails the build, so the loops cannot reintroduce it. That suite is now 114 checks (see the evals case study).

What changed

The site went from scaffold to 991 live URLs, with a median body length of 2,548 words and zero near-duplicate pages, measured by 8-word shingle overlap across 21,430 page pairs.

A full rewrite pass over 172 architecture pages and 204 short-video clusters each closed as a single tracked arc, with the backlog showing every item and its result.

Skills shown

• AI agents and multi-agent orchestration
• Prompt and context engineering
• Guardrails and human-in-the-loop review
• TypeScript, React, and Node
• Shipping in reviewable increments