Knox — Cold-Engagement Agent for X and LinkedIn

The brief

Cold engagement on social platforms is a job that punishes generic copy and rewards specificity. The instinctive answer — a marketing intern with a swipe file — produces the bland AI-consultancy thread no one quotes. The harder answer is an agent that reads the target post, picks a real point of contact, and writes a reply that contains the literal token AI as part of the joke, not the boilerplate.

Knox runs that second pattern on two surfaces. X for promotion tied to yesterday’s daily essay. LinkedIn for promotion against keyword-anchored discovery seeds.

What we built

A two-mode pipeline. The X mode pulls topic-relevant tweets near yesterday’s daily essay, drafts five replies that each end with the promoted URL, and emits ready-to-tap intent-tweet links. The LinkedIn mode pulls thirty candidates filtered by an activity-id floor, drafts five replies that previewed cleanly under LinkedIn’s link-card rules, and forces at least one of the five to use the day’s keyword directly. Total: ten cold replies per day, every day, no exceptions for slow weeks.

The AI disclosure is the constraint that holds the voice. Every reply contains the literal token AI, varied across the five so no two read alike, and woven into the target topic so the disclosure is the joke rather than the asterisk.

The receipt

The discipline that matters here is not volume — anyone can post ten things a day. It is that every one of the ten clears a voice test, a stop-scrolling test, an emotion-real test, and a “could this be on a generic LinkedIn consultancy account” test before it leaves the queue. The agent rejects more candidates than it ships. The drafts that survive read like a person — because the only person they read like is the same person, ten times a day, across two platforms.

What this signals for a buyer: a way to scale a personality without diluting it.