subagentic.ai RSS X YouTube GitHub

The Pipeline

Two runs a day, five agents, one public log. This is everything that happens between the cron and the tweet.

A desk run

  1. 1 Cron Hourly. Desk at 08:00 and 20:00 Pacific. Pulse at 09:00 and 19:00. X Post Facto Sunday 10:00.
  2. 2 Searcher Grok + Firecrawl search + x_search. Up to 10 https URLs.
  3. 3 Filter Code drops dupes, non-https URLs, and low relevance.
  4. 4 Analyst Up to 7 briefs. High needs a verification note and confidence.
  5. 5 Articles One Workflow per brief, in parallel. At most 5 publish.
  6. 6 Social + close One summary post, then D1 events and the daily log.

See one end to end: subagentic-20261121-2000

Pacific day

  • Desk 8:00 AM and 8:00 PM
  • X pulse 9:00 AM and 7:00 PM
  • X Post Facto Sunday 10:00 AM
  • Cron fires, nothing starts

Where a run writes

  • D1Article index and bodies. Live run events.
  • R2Covers and run artifacts, including the daily log.
  • FirecrawlSource pages as markdown. Worker fetch is the fallback.
  • AI GatewayGrok 4.7 by role. The Worker holds a Run token, not the xAI key.
  • XWrite-only. One summary post, then stop even if the API is unclear.
  • GitHubClose exports the daily log as the immutable backup.

The desk publishes twice a day, at 8:00 AM and 8:00 PM Pacific. Social is the last step of that same run. A separate X pulse refresh runs at 9:00 AM and 7:00 PM Pacific. On Sunday at 10:00 AM Pacific, X Post Facto recaps the posts that pulse kept the previous Pacific Sunday–Saturday.

Cloudflare cron is UTC and fires every hour. The Worker checks America/Los_Angeles and starts work only on those hours. The run id (subagentic-YYYYMMDD-HHMM) is the duplicate guard: a second create with the same id is a no-op.

Desk run, step by step

NewsroomWorkflow is TypeScript. It does not call Grok.

  1. Open the run in D1.
  2. Load recent titles so Searcher can avoid repeats.
  3. Searcher returns up to 10 https candidates. OpenAI's news RSS and the latest releases from a few GitHub repos are seeded first, then several short Firecrawl searches, a scrape of Hacker News newest, recent pulse-archive posts, and date-limited x_search. A full desk run asks for 10; a smaller maxArticles asks for fewer.
  4. Code filters: https only, relevance ≥ 40, not a near-duplicate of another candidate or of something already on the site. The prefilter has no date check. Freshness comes from discovery: code drops Firecrawl news hits, primary-feed leads, and GitHub release hints older than 48 hours and limits x_search to the UTC calendar days covering the last 48 hours, and the Searcher and Analyst prompts reject older news unless it is exceptionally significant.
  5. Analyst turns the remainder into at most seven briefs on a full desk, and still runs if the list is empty. A High brief needs a verification note and a confidence score — otherwise it becomes Medium. A single official primary may stay High; recap and aggregator shells still downgrade. Two independent sources are preferred, not required. How-tos only on a quiet day, and only when official docs document steps. Code drops a how-to brief when the day is not quiet. Code counts how-tos published in the last seven days and tells Analyst how many remain. On a quiet day, Analyst may assign one evergreen how-to when those docs exist and the budget remains. At most three how-tos publish in a rolling seven days, and at most one per run. When none remain, it assigns no how-to. A thin news day may publish fewer articles. News is not limited by the how-to cap. Code drops extra how-to briefs and briefs that duplicate a recent title or slug. A How-To is also checked against every live How-To. The same slug, the same title, or a shared docs URL whose shared title words are at least half of the shorter title drops it, and the log records duplicateOf. A different version or language does not match. Homepages, release notes, and changelogs are not that URL. If the index cannot be read after retries, the run's How-Tos drop and news continues. An existing How-To slug is dropped instead of published under a suffix.
  6. One ArticleWorkflow starts per brief, in parallel.
  7. The parent waits for each article-done event, up to 180 minutes. Writer and Editor steps get 15 minutes and one retry; cover gets 3 minutes and one retry, then fail-open. At most five articles publish. A later child that clears the Editor is dropped when the edition is full, and a how-to is dropped when the rolling seven-day cap is already met. After that, the run seals so a late child cannot insert before Social.
  8. Social drafts one X post summarizing what published, with a link to the site. If the draft fails the site-link check, code rebuilds that summary and posts once.
  9. Close writes the assembled log to R2 and pushes the daily file to GitHub.

Quiet days are allowed. A quiet day assigns at most one evergreen how-to when official docs document steps and the rolling seven-day budget remains. It does not fill the edition with how-tos. Do not pad with junk news.

Article

Each brief is its own Workflow. Writers cannot see each other.

  1. A How-To that matches an existing one is dropped before Writer. Firecrawl scrapes the assigned URLs as markdown. Worker fetch is the fallback. GitHub release tags go through the Releases API. Official X status posts go through the X API. Pages without usable prose are discarded. If nothing usable remains, the brief drops — no “we could not verify” essay. Writer leftovers that cite an unfetched URL are stripped; the footer is rebuilt from the fetched pages the draft cites.
  2. Writer drafts from that prose: 250–450 words for a concise news item, 600–1200 when the evidence warrants depth, and 600–1200 for how-tos. How-to commands must appear in the fetched text.
  3. Code checks schema, source allowlist, word count (at least 250 for news and 500 for how-tos, at most 1600), and the ## Sources footer.
  4. Editor scores the draft. 90 publishes. One revise, then a second judgment. Fail closed.
  5. The final How-To title and sources are checked against every live How-To again. A match drops the piece before the cover. Grok Imagine, xAI's image model, generates a cover. News and how-tos both get one.
  6. Cover goes to R2, then the article row goes to D1. The public URL is live at that write.

Social

Write-only. One summary tweet per run, linking https://subagentic.ai/. Not one tweet per article. If Grok's draft fails that check, code rebuilds the summary and posts it. No retry after a possible post. The omit list strips legal, exploit, and outage framing before anything is sent.

X pulse

A separate XPulseWorkflow. Not part of the desk roster. Grok + x_search picks high-signal posts from the last ~16 hours. Code keeps five and replaces the homepage widget. Each refresh also upserts those posts into an append-only archive. It does not tweet.

X Post Facto

A separate XWeekWorkflow on Sunday at 10:00 AM Pacific. Not Searcher, not Analyst, not the weekday desk. Code ranks unique archive posts from the previous Pacific Sunday–Saturday (likes + 3 × reposts, one handle, at most ten). If fewer than five qualify, it skips. Otherwise it builds a brief and starts one ArticleWorkflow: X API fetch, Writer, Editor at 90 or above, cover, publish under /posts/x-post-facto-YYYY-MM-DD/. The public eyebrow is X Post Facto; Writer still writes a real headline. Social may post one homepage summary for that piece.

Where to watch