“Coding agent” is an accurate name and a misleading one. Claude Code and Codex were built to work in a repository, and that is what they are best known for. But what they actually have is a terminal on your machine: files, scripts, package managers, exports from every tool you use, and a browser when a command-line tool isn’t enough. Most of the work in a company is reachable from there.
The question is not whether the agent can draft the deck. It can. The question is how the person who needs the deck asks for it, without a terminal, without your laptop, and without you as the relay.
Six things people ask agents for that aren’t code
Turn meeting notes into a summary. Paste the transcript or drop the notes file. The agent produces decisions, owners and open questions in the format your team already uses. A 45-minute transcript becomes a one-page summary in about a minute.
Build a deck from an outline. From a Markdown outline to a presentation file, with the company template applied, charts generated from the CSV in the folder. A twelve-slide weekly deck takes five to six minutes, most of it chart generation.
Export design assets. Pull the latest icons from the design tool’s API, export at three sizes, optimise, and drop them in the folder the site build reads from.
Draft campaign copy. Five subject lines and three body variants from a brief, in the brand voice documented in the repository, with the character limits checked.
Pull the weekly ops report. From the CRM export in the shared folder to a table of pipeline movement, with the same query as last week so the numbers line up. Under a minute once the query exists; the first week is spent writing it.
Turn a screenshot into a bug report. A photo of the broken screen becomes a reproducible issue with the relevant file paths, because the agent can read the code the screen came from.
None of these need a new tool. They need someone to ask, and a place for the answer to land.
A coding agent reaches files, exports and APIs, templates, scripts and a browser; most office work lives in those five boxes.
Why a room changes who can ask
The person who wants the deck is rarely the person with the agent. In a room, that stops mattering.
- They install nothing. They open the room in a browser or on their phone and mention the agent.
- They see it work. Each tool call streams into the room, so “is it doing anything?” is answered without a message to you.
- They get the artifact. Files and images the agent produces appear in the room’s outputs, ready to download. An image the agent generated outside the working directory still shows up.
- You stay in charge. Their request runs on your machine under the permission mode you set for them, and anything that spends money or touches a credential pauses for you.
Attachments matter more here than in coding. A screenshot is the most common way a non-engineer describes what they want. Rooms treat images and files as first-class input to the agent.
The scenario table
The same bet, from the other direction
Products built for offices arrive at this from the opposite side. Kylon, a hosted AI-native workspace, says its focus is companies without an engineering team: operators rather than developers, putting agents to work reading what already exists and carrying the next step forward, with sensitive actions left for a human to confirm.
Same observation, different conclusion. They build the workspace those companies lack; muti assumes the capability already sits on one laptop and the missing piece is a way for everyone else to ask. With no coding agent anywhere in the team, a workspace is the better starting point. With one good one already running, reaching it costs less than rebuilding it. See where agents live in 2026.
What to set up first
One agent per kind of work. Point notes at the folder with transcripts and templates, assets at the design exports, reports at the CRM dumps. Narrow directories keep the agent focused and the approvals rare.
Put the house style in the folder. A short STYLE.md on voice, a deck template, the SQL for last week’s report. The agent reads what is in front of it; give it the right things to read.
Choose permission modes by job, not by seniority. Reports and summaries can run in edit mode with no risk to code, because the agent’s working directory contains no code. Anything that can send, spend or log in should pause regardless.
Let people ask badly at first. The first request from a non-engineer will be vague. The agent will ask a clarifying question in the room, and the second request will be better. That is faster than a briefing document.
Frequently asked questions
Do teammates need their own subscription? No. The agent runs on the owner’s machine with the owner’s subscription. The room is how they reach it.
Can the agent use the tools we already pay for? If the tool has a command-line interface or an API and the credential is on the owner’s machine, yes. Using that credential routes an approval to the owner, not the requester.
Is this only for small teams? It is designed for small groups of people who trust each other: a founding team, a family, a client and an agency. Rooms are the unit, and you can have as many as you need.
If your best agent only ever hears from you, give it a room and see what the rest of the team asks it.
Published · Updated