Autonomous social-media publishing with a 4-agent pipeline.
A personal system that researches, drafts, publishes, and analyzes my social media across X, LinkedIn, Instagram, and Facebook. Four agents coordinate. Every daily brief and every draft is approvable by email in one click, and every correction feeds back into the prompts.
Publishing twelve posts a day across X, LinkedIn, Instagram, and Facebook is a real ops job. Source-scanning, topic curation, per-platform drafting, image generation, and randomized posting. Doing it by hand for a week is fine. Doing it every day for a year while shipping other work is not. Off-the-shelf schedulers solve the calendar. They leave the curation and drafting to the operator, which is the expensive half.
The expensive half is curation
The mechanical parts of publishing are solved. Schedulers post at set times. APIs cross-post between platforms. What no off-the-shelf tool does well is the front half: which trends today are worth posting, which are noise, which have been done already, and which need to be framed differently for X versus LinkedIn versus Instagram.
Twelve posts a day means twelve decisions a day about which topic goes where. Then twelve drafts. Then twelve image treatments. Then a posting schedule that does not look programmatic (every-two-hours is a bot pattern). Then post-publication engagement analysis, so tomorrow's decisions are informed by yesterday's data. A person doing this well spends two to three hours a day on it. On busy client days, the account goes dark.
The system had to run itself on the days the operator was heads-down elsewhere, without publishing anything embarrassing, and without going silent on the days the review window got missed.
Four agents, one email
Four agents coordinate. The research agent reads all configured sources (RSS, YouTube channels, X accounts and lists, Telegram bots, Instagram accounts, newsletters) on a schedule, extracts about 25 candidate topics, categorizes each (industry update, trend, tool release, opinion, thread bait), and selects the day's twelve. The posting agent takes each finalized topic, drafts a post tuned to platform, generates three or four variants, and generates two images per variant. The analysis agent reads post-publication engagement metrics and correlates them against topic category, variant style, and image treatment. The review agent proposes prompt mutations to the other three based on accumulated feedback.
Human interaction happens in email. Once the research agent finishes its daily run, the operator gets one email with the twelve finalized topics, the discarded candidates and their reasons, per-topic research context, and three inline actions per discarded item (schedule to tomorrow, include today anyway, discard entirely). If nothing is done inside the review window, the finalized twelve stand.
For each of the twelve, when the random-time trigger fires, the operator gets a second email with the drafts and images and a recommended default. Reply with a variant number, an image number, or an inline rewrite, and the post fires. Do nothing within ten minutes and the recommended default posts automatically. Same intent, platform-tuned framing: X gets shorter and punchier, LinkedIn gets a hook and a payoff, Instagram leads with the image.
The learning loop is the interesting part. Every discarded-topic reason feeds the research agent. Every draft rewrite feeds the posting agent. Every variant pick over the recommended default feeds the posting agent. Every explicit feedback text feeds whichever agent it is about. Snapshots save with timestamp, actor, context, and reason. The review agent reads the accumulated snapshots against each agent's current prompt files and proposes mutations as a diff. The operator accepts or rejects. Accepted diffs update the prompt and snapshot the prior version. Rejected diffs discard with no state change. Once a snapshot has been used to train, it is marked used and cannot be reused.
Engagement closes the second loop. The analysis agent reads impressions, likes, reposts, replies, and saves per post, and correlates against topic category, variant style, and image treatment. Patterns that consistently outperform become triggers for the review agent to propose mutations toward those patterns. Every mutation is reversible.
It runs itself now
Agent X publishes daily across four platforms. The operator's part of the job dropped from several hours a day of curation, drafting, and posting to about ten minutes of email approvals on a good day. On days the review window gets missed, the recommended defaults keep the account publishing without gaps.
Built once. Runs itself now. The prompts get sharper every week from real engagement data, not from anyone writing new prompt drafts by hand.
Every external call to X, LinkedIn, Instagram, and Facebook is wrapped in retries and fallbacks. If Instagram is down, the post still fires to the other three and the Instagram queue picks up the retry on the next window. Randomized timing spreads posts across the day with real variance. Not every two hours. Not the same slots every day. Enough variance that the pattern reads human.
Every action from every agent writes an audit row. Every human approval writes an audit row. The full history of what was posted, why, and what the prompts looked like at post time is queryable. When engagement shifts, the correlation is not a guess. It maps to specific prompt versions on specific dates that the human accepted for specific reasons.
The pattern is the same one the command center uses. Different application: content ops instead of internal ops. Same primitives: specialist agents behind one interface, human corrections as training snapshots, prompt mutations as reviewable diffs, and every state change audit-logged.
Anas