×agent-back-office.case
REDACTEDreal work, anonymized — category + problem naming · relative metrics only

Applied agents · operations

Running the back office with AI agents

Marketing, content, ads, follow-ups — a small company's whole back office, restructured as a crew of specialised agents with a human at the desk.

role · Architect & operator — multi-agent operations systemstatus · Operating internally; patterns battle-tested

§01The problem

A small product company needs the output of six departments with the headcount of one. Content calendars, ad campaigns, market monitoring, follow-up sequences — each is a part-time job, and the founder-shaped person doing all of them context-switches until nothing gets done well.

Generic 'AI assistant' usage doesn't fix this: a chat window is a very smart intern with amnesia. The work needs roles, memory, standards, and handoffs — the things that make a department a department.

[ interactive diagram loading… — six part-time jobs, one human, zero leverage ]
fig — six part-time jobs, one human, zero leverage

§02The thinking

The design borrows from management, not engineering: define narrow roles with clear briefs (research, copy, creative, distribution), give each agent standing instructions and examples of 'good', and route work through review gates where a human approves before anything external happens.

The interesting product problems were coordination problems: what does one agent hand the next, in what format, with what quality bar? The answer was artifact-shaped: every agent produces a named, inspectable artifact — a brief, a draft, a creative set — that the next stage consumes. Artifacts make the invisible pipeline auditable.

The human's job compresses into two verbs: brief and approve. Everything between is delegated — but nothing crosses the company boundary without human sign-off, because agents amplify judgment; they don't replace accountability.

+ Role-specialised agents with standing briefs

one general assistant prompted ad hoc

# consistency comes from constraints; a role with examples of 'good' outperforms a genius with amnesia.

+ Artifact handoffs between stages

free-form agent chatter

# named artifacts make quality inspectable and failures debuggable — you can see exactly which desk dropped the ball.

+ Human approval at the boundary

end-to-end autonomy

# external actions carry reputation; the approval gate costs seconds and prevents the one disaster that undoes months.

§03What shipped

The system runs as a pipeline of desks: a research agent produces market briefs, content agents turn briefs into drafts and creatives, a distribution agent packages and schedules — and every externally visible artifact pauses at a human approval gate. One person operates what previously required a team, without the output reading like a machine made it.

[ interactive demo loading… — desks, artifacts, and the approval gate ]
fig — desks, artifacts, and the approval gate

§04Outcomes

One human, six desks

sustained multi-channel output without new headcount

Debuggable delegation

artifact handoffs show exactly where quality slipped

Nothing ships unseen

external actions always pass a human gate

§05Reflection

Operating this system made me a sharper judge of the agent hype cycle. Agents are leverage, not employees; the management layer — roles, artifacts, gates — is where the real product design lives. Companies that skip it get impressive demos and embarrassing incidents.

$ cat takeaways.txt

  • Design agent systems like org charts: roles, briefs, handoffs, reviews.
  • Artifacts over chatter — inspectable outputs make delegation debuggable.
  • Autonomy at the core, accountability at the boundary.