← All playbooks

Agent systems

What it looks like when AI agents staff an entire company

Playbook 001 · 12 min

We modelled a seven-department business run by 133 scoped agents under a human oversight layer. Payroll for the equivalent org chart runs to hundreds of thousands a year; this one runs on roughly $500 a month in API tokens. The interesting part isn't that it works — it's where it breaks, and what the humans are actually for.

Human oversight

approve · escalate · own the judgment calls

Sales

24 agents

Marketing

21 agents

Support

28 agents

Finance

14 agents

Ops

22 agents

HR

9 agents

Engineering

15 agents

The architecture: one oversight layer, seven departments, one shared audit log.

The org chart, inverted

A traditional company assigns people to functions and hopes the handoffs hold. An agent company starts from the handoffs. Every department is defined by what it receives, what it produces, and who it escalates to — and each agent inside it is scoped to one narrow, testable job: read the carrier email, draft the reply, reconcile the ledger line. No agent has a job description a human would recognize; together they cover one.

The 133 number sounds dramatic, but it's really just decomposition. Sales runs 24 agents because prospecting decomposes into research, verification, brief-writing, sequencing, and follow-up — and each is more reliable as a separate agent with a separate test suite than as one clever generalist.

What each department actually runs

Support carries the biggest headcount (28) because volume lives there: triage, drafting, order lookups, escalation summaries. Finance is smaller (14) but the most tightly scoped — matching, flagging, and reporting, with every action logged and reversible. Marketing's 21 agents feed one content pipeline with compliance checks before anything ships. Ops, HR, and Engineering follow the same pattern: many small agents, one shared memory, one audit log.

Where it breaks

Three places, reliably. First, ambiguity at the boundaries — when a ticket is half support and half finance, agents ping-pong until a routing rule settles it. Second, novel situations: anything without precedent in the shared memory gets escalated, and if escalation is slow the whole lane stalls. Third, silent drift — an upstream system changes its format and four agents downstream quietly degrade. Monitoring catches it, but only because every action is logged against an expectation.

The load-bearing wall is human

The oversight layer isn't ceremony. Humans own the judgment calls: pricing exceptions, unhappy customers, anything with legal weight, anything novel. The system's job is to make those calls rare, well-briefed, and reversible. That's the real economics — not replacing seven departments, but concentrating the humans where judgment actually pays.

Under the hood

InboxCRMERPCalendar
Agent mesh · 133 scoped agents
Shared memoryAudit logEscalation queue

system design: sources in, one agent mesh, one audit trail out

The workflow, end to end

OVERSIGHT humans · policies · limits SALES DEPT 28 agents OPERATIONS 41 agents FINANCE 22 agents SUPPORT 19 agents ONE AUDIT LOG every action, reviewable policies flow down actions flow in

Four of the seven departments shown — one oversight layer, one shared audit log.

The takeaway

Agents don't replace the org chart. They compress everything that was never really a judgment call — and make the remaining human work visible, briefed, and worth the salary.

Want this system
running in your business?