Content Pipeline as a Team

A blog post, a newsletter, a report — finishing one piece of content usually means doing research, drafting, and review separately. Hand all three to a single agent and the tool list balloons, roles blur, and the output gets muddy.

This recipe splits the work across three agents and lets you drop in just a topic. The team carries it from outline → draft → review → final on its own.

What you get

You only post to the team chat once.


[You]: Write a piece on this topic — "Laptop buying guide for office workers in their 30s"

The team takes it from there.

The Planner picks up the topic, gathers sources, and lays out an outline.
The Writer drafts the piece. The final version you’ll see lives in a canvas; a copy of the draft body is also saved to the team-wide note so the Reviewer can read it.
The Reviewer reads the draft from the team-wide note and posts feedback on gaps, overclaims, and flow back to the team chat.
The Writer applies the feedback, polishes the final version in the canvas, and updates the team-wide note.

You end up with the work happening live in the team chat and a single canvas holding the finished piece — and a push notification when it’s done.

The team — who’s who?


  [You]
    │ "Write something on this topic"
    ▼
  [Coordinator]   ← orchestrates the whole team. Receives the topic and routes work.
    │
    ├──▶ [Planner]    Web search → outline
    ├──▶ [Writer]     Canvas (final) + team-wide note (draft for the Reviewer)
    └──▶ [Reviewer]   Reads the draft from the team-wide note → team chat feedback

Role	Core tools	Why split this way
Coordinator	Team chat, mentions, task tracking, schedule	Conducts the three members and consolidates progress in one place for you
Planner	Web search, page fetch	Focused on sourcing. Doesn’t write copy, so accuracy stays sharp
Writer	Canvas, team-wide note, file read/write	Focused on writing and editing. The canvas is what the user sees; the same draft body is also saved into the team-wide note so the Reviewer can read it
Reviewer	Team-wide note (read), team chat	Creates nothing new; only critiques what’s there. Keeps a critic’s stance

Before you start — create a team

This recipe only works inside a non-default team. Team-member agents cannot be created in the default team.

Click the + button at the top of the sidebar to open the Father agent, then ask:

Create a new team called “Content Team”.
The team appears in the Teams section of the left sidebar.
Create a Coordinator agent (e.g. “Content Coordinator”) and place it in that team — or ask the Father agent to set up the whole thing at once.

Build it — one natural-language ask

The fastest path: open the Father agent via the + button at the top of the sidebar and just ask in natural language. Only the Father agent can create team members and equip their tools — regular team members (including the Coordinator) cannot create other members.

Build me a 3-person content team — a Planner, a Writer, and a Reviewer.

Planner: researches the topic on the web and produces an outline. Tools: web search, web page fetch.

Writer: takes the Planner’s outline and writes the piece. Keep the version the user sees in a canvas, and save the same draft body into the team-wide note (e.g. under keys draft_v1, draft_final) so the Reviewer can read it. Apply the Reviewer’s feedback to ship the final. Tools: canvas, team-wide note, file read/write.

Reviewer: reads the draft the Writer saved in the team-wide note and gives feedback on facts, prose, and structure via team chat. Tools: team-wide note (read), team chat. Does not create new material.

Give each member a clean role description as well.

The Father agent creates the three members and equips their tools for you. Before creating each one, it checks the existing team roster — if a member already covering that role is found, it delegates to that agent rather than spawning a duplicate.

How a single ask unfolds

Once the team is in place, you only talk to the Coordinator.


[You → Coordinator]: Write a piece on "Laptop buying guide for office workers in their 30s".

[Coordinator]: @Planner please research and outline this topic — "Laptop buying guide for office workers in their 30s"
               (sent with "response required" so it knows to report back)

[Planner]:     (works quietly) gathers price ranges, CPU trends, real reviews → outlines →
               "@Coordinator research and outline are ready. Key points are ..."
               (a thread reply that also @mentions the Coordinator — wakes the Coordinator immediately)

[Coordinator]: @Writer please draft this outline — (outline attached)

[Writer]:      puts the draft in a canvas and saves the same body to the team-wide note under `draft_v1` →
               "@Coordinator first draft is ready. The body is in the team-wide note under `draft_v1`."

[Coordinator]: @Reviewer please review the draft saved by the Writer in the team-wide note under `draft_v1`.

[Reviewer]:    reads `draft_v1`, posts feedback to team chat →
               "@Coordinator three things: 1) ... 2) ... 3) ..."

[Coordinator]: @Writer please apply this feedback and finalize — (feedback attached)

[Writer]:      updates the canvas and saves the final body under `draft_final` →
               "@Coordinator final version is ready."

[Coordinator → You]: "The piece is ready — check the canvas."

The point is that everything happens in one team chat, so you can follow who is where at a glance.

Where the result shows up

Where	What you see
Team chat	All messages between Planner, Writer, and Reviewer in chronological order.
Canvas	The user-facing finished piece, written by the Writer. Draft → revision → final all live on the same page.
Team-wide note	Where the Writer leaves the draft body so the Reviewer can read it. You can inspect the entries yourself in the Team State panel on the right side of the team chat.
Push notification	When the piece is done, the Coordinator pings you (must be enabled in `Settings > Notifications`).

Polishing — make it yours

After one full run, tune it to your workflow.

Add an SEO specialist: insert another member that produces keyword analysis and title candidates before the Writer step.
Tighten the Reviewer: bake a checklist into the Reviewer’s role description — “every claim needs a source”, “remove superlatives and filler”.
Share a style guide: company tone, banned words, voice — store them in a team-wide note so all three members read the same rules every time (see Advanced below).
Run on a schedule: give the Coordinator the schedule tool and set a recurring time — it will produce one piece automatically on each run.

Advanced

The section below is not needed for normal use. Read it only when the team’s flow does not behave as expected, or when you want to package this team for reuse.

Mention model — who wakes whom?

When a team is created, every member is connected to every other member with a bidirectional link (mesh). In the default state this means any team member can wake any other member by mentioning them — the Coordinator, the Planner, the Writer, and the Reviewer can all mention each other freely.

You can prune specific links in the team page’s Graph tab. A mention going in a pruned direction still appears in the team chat as text, but no run is enqueued — the recipient will see it on their own schedule.

This recipe routes everything through the Coordinator, but you can configure members to hand work directly between themselves — the default mesh already has every path open.

response_required option: when the Coordinator sends a mention with this option, the mentioned member is guided to mention the Coordinator back once the work is done. This option only works when a mention path is open to the target. It has no effect if the link to that member has been pruned.

Thread replies: when a member sends a report back (e.g. “@Coordinator all done”), they must include both in_reply_to (the mention ID from the Coordinator’s original message) and @Coordinator in the message body. Both are required — in_reply_to unlocks the path back to the original sender even when no edge exists, and @Coordinator is what actually triggers the wake-up. Without the @mention the reply just sits in chat.

Store the shared guideline in the team-wide note

A style guide, banned words, or company tone — anything every member must read every time — should not be saved in the Coordinator’s personal memory. Personal memory is private to that agent.

Instead, ask the Coordinator once:

Save our team’s shared style guide into the team-wide note under the key style_guide. Content: …

The Coordinator writes this to the team-wide note (Team State). From then on, the Planner, Writer, and Reviewer all read the same style_guide key when they work, producing consistent output. Writing to the same key again overwrites the current value — previous content is not retained. If you want to keep an older version, save it under a different key.

Only the last writer of a key, or a Father agent, can delete an entry. The Father agent is the separate system-level agent you open via the + button at the top of the sidebar — it is not the same as the Coordinator role created in this recipe. Other members must ask the last writer to remove it.

You can also inspect the current entries yourself in the Team State panel on the right side of the team chat.

Reuse it via a team preset

Once you like the setup, you can publish it as a Team Preset to the marketplace so you (or others) can install the same team again later.

On the team chat screen (/agents/team/<team>/chat), click the Share button at the top right (not available in the default team).
The team structure (agents, connections, tools, skills) is packaged as a single bundle at publish time.
Anyone who installs the preset from the marketplace gets an identical team cloned into their account.

Edits made to your original team after publishing do not flow back into the existing preset. Republish to ship a new version.

A note on cost

All LLM and tool calls inside the team draw from your credits as the team owner.
More members and more active tools mean more cost per piece. A read-only role like the Reviewer can usually drop to a lighter model tier with no real drop in quality.
If you plan to run this on a schedule, do one or two manual runs first to estimate the cost per piece.