The agent team and custom agents

See what each agent did, change its settings, pause the team, or add your own agents.

Each brand has a team of agents that runs the pipeline: strategist, researcher, writer, SEO, brand editor, designer and publisher, with the AI assistant as coordinator and an analyst that learns from your reviews.

The Team page

Team shows each run's progress, how the team is improving (for example the first-pass rate) and a log of each step. You can:

  • give the writer standing instructions;
  • give the editor a brand checklist, one point per line (for example "name the warranty period"). The writer sees it, and the editor checks every draft against it before review: a draft that misses a point goes back to the writer, and if it still misses it after the last return, Review shows a warning;
  • switch the researcher or designer off, or keep the researcher to your knowledge base only (by default it also searches the public web for sourced statistics, standards and market data, which show under "Web sources" on Review; ad copy, video scripts, product copy and Q&A answers never use the web, only your own knowledge and products);
  • restore an earlier version of a member's settings;
  • pause the team, which stops new batches.

Custom agents

Build your own agents from a template, check them with a trial run, and place them in the brand's flow:

  • Before the writer: adds notes to the writer's prompt.
  • After the editor: a check that can send a draft back (1 to 3 times) or add a warning.
  • Weekly: a summary and topic ideas you can add as topics.
  • On an event: a batch is ready, something was published, a publish failed, a new knowledge document arrived, you sent a piece back on Review, or a batch paused because the month's topics ran out.
  • Instead of the researcher: its notes replace the researcher's public web search. Your knowledge base is still read.
  • Instead of the writer: it writes each version with its own instructions and the material it may read. Instead of the team's writing playbook and your writer instructions, the writer follows the agent's instructions. Everything else stays the same: the hard rules, the editor's checks and your review. Its trial run writes the latest piece in review again, without saving it. While it writes every piece, the analyst only proposes changes to the editor's checklist or returns, since there are no writer instructions left to improve.
  • Instead of the editor's voice check: it checks each draft's tone and key messages in its own way. What it finds shows on Review as a warning; the draft is never sent back for it, and the editor's other checks stay. If it can't run, the editor's own voice check runs.
  • Instead of the designer's picture: it reads each draft and describes the photo for its cover. Finding a matching library image, drawing the AI image, the image checks and your cover template work as before. It doesn't run while the designer is switched off on the Team page.
  • Instead of the strategist: before each batch, when planned topics are short of the batch size, it picks which free keywords become topics and writes their titles by its own instructions. Each keyword still goes to one topic, a title always contains its keyword, and any places it leaves open the strategist fills. It has no conditions, and only one agent can do this. Its trial run plans 5 topics from your free keywords without saving them.

Order, running at the same time, and conditions

On the Team flow page you set each agent's position. Agents in the same place run in that order:

  • One after another: a later agent before the writer reads the notes of the ones before it. After the editor, a check that finds a problem sends the draft back straight away, and the checks after it wait for the next draft.
  • At the same time: tick "At the same time" to run an agent alongside the one before it. This is quicker, but neither sees the other's work.

Under Only for some pieces, tick the platforms, languages or content types an agent works on. For example, a stand-in writer can be used only for YouTube and TikTok, while the team's writer keeps the rest. If you tick nothing, the agent works on every piece. Only one agent can stand in for the same member on the same pieces.

A check after the editor can also look at the draft itself: under Only on some drafts, list words or phrases (one per line), tick products from your catalogue, or both. The check then runs only on drafts that mention one of the words (whole words, any case) and, if you ticked products, name one of those products. You can also give a length (at least and / or at most so many words in the body; Chinese and Japanese count characters) and whether the draft states a price (an amount with a currency) or states none. Every item you fill in must match. For example, a compliance check can run only on drafts that mention "warranty", and a price check only on drafts that state a price. Up to 20 words and 20 products.

Custom agents only read the knowledge documents and web addresses you tick for them, plus the brand data their tools allow: the next 30 days of the calendar, the last 90 days' results, matching images in your library, and the keyword data you already have (no new paid lookups). Trial runs and weekly runs use run credits; steps inside a batch count toward the topic.

With the Calendar events or AI images tool, a weekly or event agent can also propose events for your calendar and images for your image library. They wait under From your agents on the Team page until you confirm each one: a confirmed event goes on the Calendar page as your own event, and a confirmed image uses one AI image from the month's allowance (on Free, once they're used up the proposal waits for next month). A trial run only shows what the agent would propose.

The analyst's reviews (and the test drafts it writes to check a proposed change) also use run credits. Automatic reviews only run while more than 20% of the month's included run credits are left, so they don't use up what your custom agents need.

More in Generation and review