Your Paragraph agent is an orchestrator: it reads what you asked for and hands the job to the sub-agent built for it. Publishing isn't one job. Drafting an essay, cutting a video, and designing a website need different tools, different craft, and different judgment. We built your agent as a fleet of nine so the one doing your work is the one built for it.
Writing: Drafts and edits posts, newsletters, and essays. It edits already-published pieces without flattening the buttons and embeds in them, and repurposes one piece into another format.
Publishing: Runs the publish flow. The pre-publish review, readiness recommendations, the slug, whether it goes to your list, test emails, and your subscribers, including a CSV import.
Media: Generates and edits images: post covers, social cards, illustrations. It also resizes, crops, and converts files already in your library.
Video: Builds real videos, with a narrated voiceover and a music bed by default. Audio is generated first so the scenes are cut to it. You watch a live preview, then export an mp4.
Website: Builds, restyles, validates, and deploys your website. It owns the theme gallery and fixes its own build errors before you see them.
Social: Owns X and LinkedIn. It drafts posts, threads, and X Articles for your review, publishes the ones you approve, and reads your timeline to see what landed.
Research: Looks things up and analyzes. Web research, your analytics, your own back catalog. It's read-only, so it never drafts, posts, or changes anything.
Scheduling: Owns work that happens later. One-off scheduled tasks, recurring workflows, follow-ups keyed to publishing, and your subscriber lifecycle emails.
Brand: Owns your visual identity: colors, type, spacing, and reference images. It reads a site's real stylesheets rather than guessing from a screenshot, then renders a mood board so you react to real output.
Interacting with the fleet is entirely conversational. You never have to pick a sub-agent or learn special commands; you simply describe what you want, and the orchestrator handles the routing behind the scenes. As work progresses, a clear badge on each step names the specific agent currently running it, letting you see when a request lands with the website agent rather than the social one.
Each agent is deliberately focused by design, carrying only the tools required for its own domain so it reaches for the right capability rather than choosing from everything at once. Despite this specialization, they all share a single memory containing your voice, goals, and notes. This ensures that passing work seamlessly between different agents never causes your style to drift.
Sequential multi-step requests, such as drafting a post and tweeting a link to it, execute in a precise order. The writing agent completes the draft first, and then the social agent picks up the newly created post to handle the rest of the workflow.
The badges appear while the fleet works. If you reload the page mid-task, finished steps collapse into one row and the labels go with them. Some agents need a connected account to finish the job: the social one needs X or LinkedIn connected before it can post what you approve.
We're building an owned engine for your companies and founders. Ask your agent for something that needs real work, a website or a video, and watch which one takes it. See what else it can run for you at paragraph.com.
Colin Armstrong
