All posts

3 min readScenariosWork

A marketing team’s week with one coding agent and a room

Five people, one founder’s Claude Code, one room. Monday’s deck, Tuesday’s campaign variants, Wednesday’s landing-page copy, Thursday’s report from the CRM export, Friday’s screenshots turned into bug reports. What they asked, what ran, what paused.

By Noah· Product at Muti AI

A coding agent is an operator for anything reachable from a terminal, and most marketing work is: files, exports, templates, a browser when a command-line tool isn’t enough. What a marketing team lacks is not the capability but a way to ask for it. This is one team’s week, five people and one founder’s Claude Code, reached through a room. Names are illustrative; the requests are the kind we see.

Setup, once

The founder created an agent called growth, pinned to a folder with three things in it: a STYLE.md describing the brand voice and banned words, the deck template, and a reports/ directory where the CRM export lands every Monday. Permission mode acceptEdits, so the agent writes files without asking and commands still ask. The five marketers got a standing grant. Anything that sends, spends or logs in pauses for the founder.

Monday: the deck

Request (Priya, on her phone): “Monday’s pipeline deck from outline.md, template applied, charts from this week’s export.”

What ran: the agent read the outline, generated four charts from the CSV with the same queries as last week, filled the template, and posted a .pptx to the room. 6 minutes. Nothing paused.

What changed the following week: Priya added “compare to last month” to the outline file and never asked again.

Tuesday: the campaign

Request (Dan): “Three subject-line variants and two body variants for the launch email, brand voice, under 50 characters for subjects.”

What ran: drafts in the room in under a minute, with character counts and a note on which banned word it avoided. Dan asked for a fourth variant “more direct”; the agent produced it in the same thread.

What paused: nothing. Drafting doesn’t send. Sending would.

Wednesday: landing copy and a translation

Request (Mei): “Rewrite the hero and three feature blurbs for the new landing page in brand voice; then a Simplified Chinese version.”

What ran: the copy, then the translation, then, because the agent could see the site repository in a sibling folder was not in its working directory, a note saying it had written the copy to landing-copy.md for an engineer to apply.

Why that matters: the agent’s working directory is the boundary. The marketing agent cannot touch the site’s code, by design. See permission modes, explained.

Thursday: the report

Request (Priya): “This week’s pipeline table and the month-to-date chart.”

What ran: the same SQL as last week against the new export, a table in the room, a chart image. 40 seconds.

What paused: the agent proposed calling a paid enrichment API to fill missing company sizes. That is spend, so it paused for the founder, who rejected it with a note: “not this week”. The report went out without the column.

Friday: screenshots into issues

Request (Sam, four screenshots): “These look broken on mobile. Write them up.”

What ran: four issue drafts, each with the likely component named because the agent had the design system docs in its folder, and a severity guess. Sam pasted them into the tracker.

What paused: nothing. If the agent had been in the site repository it could have opened the issues itself; that agent belongs to engineering, in another room.

Requests, outputs and pauses across one week.Requests, outputs and pauses across one week.

What the week shows

  • Nobody installed anything. Five people used a browser and, mostly, a phone.
  • The house style lived in a file. STYLE.md did more for consistency than any briefing.
  • Boundaries were directories. Marketing’s agent saw marketing’s folder. Engineering’s agent saw the repository.
  • One thing paused all week, and it was the one that would have cost money.
  • The agent got better as the folder did. Every request that recurred became a file the agent could read next time.

Frequently asked questions

Did the marketers need to know Markdown or the terminal? No. They typed requests the way they would to a colleague. The agent asked one clarifying question all week.

What did it cost? The founder’s existing Claude Code subscription. Per-request usage is recorded per person, so the founder can see it. See borrowing an agent.

Could the agent have sent the email? Only if the founder approved. Sending is a spend/credential action and pauses every time.

Your team already has the questions. Give your agent a room so they can ask.

Published

Give your agent a room.

muti is free during the private beta. Connect the machine where your agents live and invite the people you trust.

Get started

Keep reading

6 min readComparison

QoderWake and muti

Of every agent product launched in 2026, Alibaba's QoderWake is the closest in shape to muti. Both run on your own machine, and both let a group send it work, through a chat that has to be paired first. They split on whose agent it is, what it runs, and who approves when a colleague's request reaches a delete.

9 min readComparison

WorkBuddy, QwenWork and QoderWake

Tencent's WorkBuddy lets you drive the agent on your own computer from WeChat and WeCom. Alibaba's QwenWork merges desktop, cloud and DingTalk agents into one product. QoderWake runs a team of digital employees on your machine behind a single bot the whole group can mention. All three let colleagues use the agent. They answer how in very different ways.

5 min readPersonal agents

An agent with its own phone number

Manus Cue gives each personal agent an email address, a phone number, a wallet and a computer. That makes the agent reachable by anyone who has the number, which changes the permission question from "what may my agent do" to "what may a stranger make it do". What Manus 2.0 shipped, and what reachability asks of an agent.