Blog pipeline as an agent production line (state in GitHub Issue comments)

A multi-stage agent pipeline for writing public blog posts: Research → Questions → Writing → Critique → Draft → Deploy. State persists as comments on a single GitHub Issue per article; each stage's agent appends a comment with its output and a step number. The next stage's agent picks up from the latest comment. The shipped artifact is a public blog with ~3K visitors/month.

"Agents forget" is the universal failure mode of multi-step content pipelines. Treating a GitHub Issue as the durable memory makes the pipeline observable (the Issue is the audit trail), restartable (any stage can re-read all prior comments), and human-pausable (a human can append guidance at any step).

1. Define the pipeline skeleton as JSON: list of stages, expected inputs/outputs per stage, agent prompt path per stage.
2. Trigger: when a lecture transcript hits the inbox folder (or any source event), open a new GitHub Issue with the topic in the title.
3. Stage 1 (Research): agent reads the trigger context, runs web research, appends a comment with findings and step: research.
4. Stage 2 (Questions): agent reads all prior comments, drafts clarifying questions, appends.
5. Stages 3–5 (Writing, Critique, Draft): same pattern. Each agent reads everything below it.
6. Stage 6 (Deploy): publish to the public blog, append the live URL as the final comment, close the Issue.
7. Each agent's instructions explicitly say: "read every prior comment, then append yours — do not rely on conversation memory."

• Don't let agents edit prior comments. Append-only is what makes the audit trail trustworthy; once edits are allowed, an LLM will silently rewrite a step-2 finding to match a step-4 conclusion.