WorkAgnt AI

New X's algorithm is out and this agent knows it. Just paste your post, the agent will analyze it, and recommend changes if needed. But not just that: WorkAgnt now lets you create AI employees that understand X's algorithm based on the open-source codebase. You've probably heard the usual advice: "Post at 9am", "Use threads", "Engage more", but have you ever wondered what's actually happening under the hood when X decides to show your post in someone's for your feed? X open-sourced their recommendation algorithm. We read the entire codebase. Then we built an AI employee that understands it. What the Algorithm Actually Does Your post goes through a pipeline called Home Mixer. It pulls candidates from two sources: Thunder: an in-memory store that tracks posts from accounts you follow. Sub-millisecond lookups. This is your "in-network" distribution to the people who already follow you. Phoenix: the ML ranking engine. A Grok-based transformer model that scores every candidate post. This is how your post reaches people who don't follow you. The "out-of-network" discovery. Phoenix uses candidate isolation, each post is scored independently, not relative to other posts in your feed. This means your post's score is consistent and catcheable. It either scores well or it doesn't. What Signals Actually Matter The scoring model predicts engagement probability. Here's the hierarchy, from highest impact to lowest: Replies - the strongest signal. A post that generates conversation ranks dramatically higher. Retweets/Reposts - signals that content is worth sharing. Likes - low friction, lower signal value. Bookmarks - indicates utility or reference value. Dwell time - how long someone pauses on your post. Negative signals are powerful. A single "Not Interested", mute, or block doesn't just affect that one user, it trains the model against distributing similar content. What Kills Your Reach Based on the algorithm's actual code: External links: the algorithm de-prioritizes posts that just link out Excessive hashtags: pattern-matched as spam behavior Engagement bait: "Like if you agree" is detectable and penalized by quality filters Rapid-fire posting: dilutes per-post engagement, hurting individual scores Muted keywords: if your post contains terms users commonly mute, distribution drops What Works (And Why)? Threads perform well because each reply in a thread is a separate candidate. Phoenix can independently surface any reply from your thread to anyone's feed. More candidates = more chances to rank. Media attachments (images, video) get engagement boosts in the ranking model. The content understanding service (Grox) processes your media and feeds signals to Phoenix. Follower graph matters for in-network distribution. Mutual follows and frequent interactions boost how Thunder prioritizes your posts for those users. Community Notes can suppress reach. If your post gets noted as misleading, distribution drops. Try It Before You Post We built an AI employee on WorkAgnt that knows all of this. Paste your draft post, and it'll analyze it against the actual ranking signals: Hook strength (affects dwell time) Engagement potential (will it trigger replies or just likes?) Reach killers (links, hashtags, bait patterns) Format optimization (thread vs single, media impact) Specific rewrites with explanations tied to the ranking logic Try it free 👇