The agent team Ā· Happened

Nine agents planned, five built

My plan for an AI newsroom had nine agents. For the first weeks the whole team was one session wearing the editor's persona.

Still true Accurate as of
Grid of nine planned agent roles. Built, with dates: editor-in-chief April 17, writer-researcher April 20, analyst June 10, publisher June 16, social editor June 19. Never built: content strategist, SEO analyst, visual producer and monetization strategist.

The short version: in April I wrote an operating manual for a publication run by a team of nine AI agents. Five were ever built. For the first weeks the ā€œteamā€ was one Claude session that loaded the editor-in-chief’s persona and memory and became the editor. Real delegation only arrived where a task could run on its own and produce something I could check.

Get the week's seven in one email, every Monday.

What went wrong

The plan had nine roles across four phases: an editor-in-chief, a writer-researcher, a publisher, a social editor, an analyst, a content strategist, an SEO analyst, a visual producer and a monetization strategist.

In practice I worked with one session at a time, and that session loaded the editor’s persona and became the editor. The editor’s own first memory note said it plainly: ā€œthe only agent on disk… Don’t pretend to delegate.ā€ When the writer agent finally got its first assignment, it did 78 tool calls of research, twice, and couldn’t save a word, because a leftover settings file in a nested folder overrode its permissions. For several weeks the editor drafted the weekly roundups itself, because that was easier.

What we learned

  • Most ā€œmulti-agentā€ setups are one session wearing hats. That’s fine. Say so, and design for it.
  • An agent earns its place when a task can run headless with a checkable output. The writer became real when it drafted on a schedule and produced a file I could review. The roles with no such task never needed to exist.
  • Unbuilt roles get absorbed. The editor took over strategy; the analyst took over SEO. Nobody missed the separate agents.

What went right

The five roles that shipped each had a concrete job, a schedule and an output: a draft, a published page, a batch of posts, a report. In two weeks of June, the jobs running unattended went from 4 to 13.

How to do it (as of October 2026)

  1. Start with one persona, loaded into whatever session you’re in.
  2. Add an agent only when a task can run without you and leaves an artifact you can check.
  3. Keep each role’s memory as a versioned file in the repo, not in a tool’s private memory.
  4. Every month or so, check which roles actually did work. Fold the rest back in.

Status

Still true. The publication runs on those five roles today.

The week's seven, by email

Entries post here daily. Every Monday, one email with the past week's seven. Unsubscribe anytime.

Let's Connect