Hi all. We’re comparing four ways to turn one blog post into a LinkedIn post, an X thread and a 5-slide carousel: a plain cron job, cron plus one LLM call, a LangGraph graph, and an autonomous agent (Hermes Agent). Same models, same metering proxy, same gates for every arm. The design was pre-registered before any counted run.
Our LangGraph arm is a StateGraph: fetch, draft, validate (code), critique (LLM), revise, with at most 2 revisions and recursion_limit 20. It uses ChatOpenAI against an OpenAI-compatible Workers AI endpoint (llama-4-scout and deepseek-v4-flash).
- Graph code (about 100 lines): social-publish-4way-2026-10/arms/arm3_langgraph.py at main · kelvindesman/social-publish-4way-2026-10 · GitHub
- Pre-registration and deviations: social-publish-4way-2026-10/PREREGISTRATION.md at main · kelvindesman/social-publish-4way-2026-10 · GitHub
- Per-run results: https://lokerdollar.com/data/social-publish-4way-2026-10.csv
Two questions:
- Is there anything in the graph that a LangGraph user would call a strawman, such as a node or edge you’d build differently for this job?
- Is anything missing that would change publishability or cost, like structured output or a different retry policy?
We’ll quote or link your answer in the write-up, including disagreement. Thanks for looking.