Set up a marketing team that runs itself with Soot

A one-person company has the same marketing chores as a team of ten: read what the audience is talking about, write something worth reading, post it on four channels, and do it again tomorrow. The usual shortcut is one agent with one long prompt. Either it posts filler every day, or it stops at every step to ask permission, and you are back to doing the work.

This guide sets up the other shape: four Soots with separate jobs, a shared workspace, two schedules, and one place where they ask you only for what a person must do. It is the team 2found runs every day for 2found.dev, trysoot.com and 2novel.app, published as a copyable example in Soot Forge.

The team

| Soot | job | tools it is granted | | --- | --- | --- | | Editor | runs the cycle, reviews, publishes | the shared pool, your CMS, product knowledge | | Radar | reads where your audience talks, writes the daily brief | web_fetch, web_search, the shared pool | | Writer | turns one angle into an article and social posts | the shared pool, product knowledge | | Poster | posts reviewed posts from your logged-in browser | browser-use, the shared pool |

They work as one Soot group. The editor is the entry: a schedule starts it, and it hands work to the others inside the group's limits (8 delegations, depth 2, 2 at a time, 120 model calls per cycle in the example).

A day looks like this:

08:00  Editor → Radar   scan Hacker News, Reddit, dev.to, your feeds → brief-2026-10-17
       Editor → Writer  one angle → article-…, post-x-…, post-linkedin-…
       Editor           review, fix, publish the article, hand posts to the Poster
       Editor → you     only "post this video by hand" or "decide this"
15:00  Radar            afternoon scan, merged into today's brief

Three places for three kinds of knowledge

The example splits what the team knows on purpose, because mixing the three is what makes agent teams drift.

• Procedure lives in skills. How to scan, how to write a brief, the review checklist, how to post: skills in one pack, written once and shared by every role.
• Your company lives in the missions. Audience, sources, channels, languages and the daily limit sit in a few lines marked Edit these lines.
• Product facts live on your website. The source points at your llms.txt; the runtime mirrors those pages into a knowledge pool, every 6 hours by default, and every Soot looks a fact up with knowledge_search before stating it. A claim your pages do not make does not get published.

The shared data pool is the team's desk. Teammates hand on document ids (brief-2026-10-17, article-agent-approvals-en, post-x-agent-approvals), never pasted drafts, and every piece carries one status tag: draft, ready, published, handed-off or discarded. A cycle starts with the leftovers of the last one, and you can open any piece in Nest.

What reaches you

The editor publishes what passes its review. It stops and asks only when a piece would announce a price, date, partnership or legal or security claim your pages do not state, answers or criticises a named person or company, or depends on a claim it cannot check.

When it asks, the question is one card in Nest's Cần bạn (needs you) list, with up to four one-click answers and the Soot's recommendation marked. It also names a safe default, such as "hold the piece", and a time. If nobody answers in that time, the runtime takes the default and the team carries on. Your silence never stalls the next cycle.

Some things are a person's job: filming a video, or posting on a channel nobody connected. Those come as one card per cycle with the exact text to paste. Reply with the post URL and the editor records it.

If you would rather approve every post, you do not need a different team. A tool can be declared "approval": "operator", so the run pauses in Nest until someone approves or refuses. The skills stay the same. The gate belongs to the host.

That older shape is still in the catalog: the `groups/marketing-team` template gates every publication on an operator approval. This guide follows the autonomous example, not that template.

Set it up

You need the Nest desktop app (download), an AI provider, and the example's soot-source/ folder.

1. Copy and rename. Replace the acme prefix in file names and contents with your own short id. Soot, group, pool and schedule ids are shared by the whole runtime, so a second team's marketing pool would read the first team's briefs.
2. Point it at your company. Edit the Edit these lines blocks in the four missions, the knowledge URL in source.json, and the schedules' time zone.
3. Install the packs the Soots ask for: research, data-pool, and, if you use them, website (a CMS) and browser-use (your logged-in browser).
4. Deploy the source. With the soot CLI:

   soot bundle --refresh ./soot-source
   soot source deploy ./soot-source

The CLI is still a preview available to people with access to the 2agent repository, and the desktop app does not install it. With Nest alone, recreate the same Soots, group and schedules by hand from these files.

1. Grant in Nest. On each Soot's profile, grant the shared data pool and the tools it may use on this computer. Create the pool, and bind your CMS credential if you publish through one.
2. Run one cycle by hand. Open the morning schedule and run it once; read the editor's report before you trust the schedule.

Without a CMS credential or a logged-in browser, nothing goes out on its own. Everything the editor would post arrives in Cần bạn as text to paste. That is a fine way to spend the first week.

What we changed after running it

The design notes in the example come from 2found's own runs:

• A schedule only names the cycle. When the schedule text repeated the procedure, the two copies drifted apart and the team followed whichever it had read last. The schedule now says "run one cycle as your mission describes".
• Spend tool calls on the output. A radar told to triage everything ran out of calls before it wrote the brief. Its skill now says that three sharp items beat a perfect list and no brief.
• Silence beats filler. A daily limit of one article and three posts, plus "nothing worth writing: say so and stop", keeps the blog from filling with weak posts.

What it costs

Soot is a preview with no published price. What you pay for is the AI provider you connect and the computer that runs it. On 2found's server, the runtime serving seven Soots, among them this team, a support chat and a feedback desk, used about 42 MiB of RAM and under 1% of a CPU core.

Start from the marketing-team example, and read orchestration for how a group delegates.