Seven Agent Files In One Repo. The Unattended Mode Loads None Of Them.

Seven agents, five layers, one repository. A strategist that reads what is climbing and writes the calendar. A scriptwriter that turns one calendar row into a full script. Thumbnail and SEO agents that package it. A production agent that renders the cut. A publisher that uploads on schedule and an analytics agent that hands the numbers back to the strategist, so next month’s calendar writes itself.

We have been through that architecture before, from the video version of the same material: a closed loop gets better at whatever number you put into it, and choosing those numbers is the real decision. This card is the still version, and it adds something the video did not — file paths. agents/strategist.md. pipeline/render.yaml. A command palette with /thumbnail-ab and /seo-pack in it. Once a diagram names files, you can check it against the documentation for the tool those files belong to.

So this post asks a different question from the last one. Not what should the loop optimise, but: on the morning nobody opens a laptop, what in that repository actually runs?

Seven agents, four kinds of thing

Start by writing down what the card actually puts under each heading. The panels are a pleasure to read and the paths are legible, which makes this easy.

None of that is wrong as a sketch. But four different mechanisms are being numbered as if they were peers, and only one of them — the subagent — is the thing most readers picture when they see the word agent.

Subagents live inside a session. This loop runs for a month.

The subagent page in the Claude Code documentation is explicit about scope. Each subagent runs in its own context window with its own system prompt and tool access, the main conversation delegates to it when a task matches its description, and — the line that matters here — subagents work within a single session. For independent sessions running in parallel, or for sessions that message each other, the page points you somewhere else entirely.

Now read the fifth layer again. The publisher uploads on a schedule. The analytics agent tracks click-through, retention and views at 48 hours. The calendar rewrites itself monthly. That is a loop whose shortest leg is two days and whose full cycle is four weeks.

A session is not four weeks long. A session is however long you sit there, or however long a scripted run takes before it exits. Nothing in agents/publisher.md wakes up on the first of the month. The file is a description of how an agent should behave once something has already started it, and the card draws no arrow into the start.

This is not a nitpick about terminology. It is the difference between a repository you clone and a system that runs. Everything inside the diagram is real; the thing that makes it happen without you is outside the diagram and is not drawn.

And the recommended unattended mode skips the folder

Here is where it gets genuinely counterintuitive, and it is worth knowing before you spend a weekend on this.

To run Claude Code without a person at the keyboard you use the non-interactive mode, claude -p, from a script or the Agent SDK. For that, the documentation recommends adding --bare, which exists to make runs reproducible across machines by skipping auto-discovery — and it lists what gets skipped: hooks, skills, custom commands, subagents, installed plugins, MCP servers, auto memory and CLAUDE.md. The page calls it the recommended mode for scripted and SDK calls, and says it will become the default for -p in a future release.

Read that against a card whose entire premise is a folder of agent files. In bare mode, agents/strategist.md is a text file nobody opens. So are the other two. The slash commands are not registered. Claude gets Bash, file reads and file edits, and nothing else, unless you hand the pieces over on the command line: --agents for the definitions, --settings, --mcp-config, --append-system-prompt.

So the seven agents can absolutely run unattended. But the thing that makes them run unattended is a line in a shell script, not a layout in a repository. “Nobody in the middle” is a flag, not a folder.

What a run with nobody present actually needs

Put the two together and the shape of the missing half becomes obvious. The card draws the middle of the chain beautifully and stops one step short at each end.

One warning about cloning a stranger’s repository

The card offers a free GitHub repo. It does not name it or link it — the caption asks you to comment to receive it — so I have not seen the repository and nothing here is a claim about it. This is about what happens to any repository you run this way.

The same documentation page carries a note that belongs in every “clone this and run it” post and is in almost none of them. Without --bare, a -p session runs the hooks in a project’s .claude/settings.json and connects the servers in its .mcp.json, even in a folder you have never trusted. A -p session shows no workspace trust dialog and no per-server approval prompt.

Interactive Claude Code asks you about a new folder. A scripted run does not. So the exact move this card invites — download somebody’s YouTube automation repo, point a scheduled script at it, go to bed — is the move that skips the question you would otherwise be asked. Read .claude/settings.json and .mcp.json before the first run, or use --bare and pass only the pieces you have read. If you want the longer version of which files in that folder do and do not load, we went through the whole map row by row against the documentation here.

What the card gets right

It would be easy to read the above as a takedown. It is not. Line by line, this is one of the more careful cards of its kind, and two of its best lines are in small grey type at the bottom.

Four honest caveats

1. I have not seen the repository. Everything above is read off the card’s own panels and checked against published documentation. If the repo ships a scheduler, a runner for render.yaml and correctly placed .claude/agents/ files, then it answers most of this and the card simply under-sells it. The gap I am describing is a gap in the diagram.
2. The documentation I checked is Claude Code’s. The card names Claude, so that is the right manual, and the subagent, skills and non-interactive pages are where the quoted behaviour comes from. If you build the same seven roles on a different stack, the specific flags change; the shape of the gap — something must start the run — does not.
3. Bare mode is a recommendation, not a requirement. You can run claude -p without --bare and keep auto-discovery, in which case the folder does load. I have led with the bare-mode behaviour because the documentation recommends it for scripted calls and says it will become the default, and because the same page explains what you trade away by not using it.
4. The missing file for agent 07 might be deliberate. The card pairs agents 06 and 07 on one panel and 03 and 04 on another, so the analytics agent may simply share the publisher’s file. I have counted what is printed, which is three subagent paths for seven agents, and I cannot tell you which it is.

The one-line version

The seven agents are real, the loop between them is the good idea, and the two things that would make it run with nobody in the middle — a scheduler to start each run, and the --agents flag to hand the definitions to a scripted session that would otherwise ignore the folder — are the two things the card does not draw.