Running a one-person company with AI agents: departments

AR Labs organises one person's work as if it were a company with five departments. Each department has a purpose, a set of activity streams (named, recurring kinds of work) and its own label in the task tracker. Today Richard M. Thompson, who founded AR Labs, fills every department role and switches between them during the week. The structure exists so that the planning system can allocate time between departments, and so that an AI agent (a language model run in a loop with tools, so it can act, not only reply) can later take over a whole role without the rest of the system changing. This page describes the structure for anyone organising work so agents can share it; our own allocations are not published.

How it works

The five departments, each with a one-line purpose:

Department	Purpose
Engineering	Build things: demos, client work, implementation
Business Development	Get work: opportunities, proposals, client relationships
Content & Marketing	Tell people: articles, building in public, the website
Product	Ship products: product strategy and requirements
R&D & Infrastructure	Improve capabilities: tooling, multi-agent systems, research, skills

Above the departments sits a three-layer model. Richard is the governance layer: he sets vision and constraints, approves pivots and is the final escalation point. The strategy agent is the orchestration layer: it reads department status and writes direction (see strategy agent planning cadences). The departments are the execution layer.

flowchart TD
    F["Richard: governance"] --> C["Strategy agent: orchestration"]
    C --> ENG["Engineering"]
    C --> BD["Business Development"]
    C --> CON["Content & Marketing"]
    C --> PRD["Product"]
    C --> RND["R&D & Infrastructure"]
    ENG -.->|"status"| C
    BD -.->|"status"| C
    CON -.->|"status"| C
    RND -.->|"status"| C

Activity streams and goal trees

An activity stream is a kind of work a department does again and again, with a stated measure. Examples from our own map: demo building and client delivery in Engineering; proposals in Business Development; content creation, web presence and video in Content & Marketing; skills development, infrastructure development and technology research in R&D.

The activity map connects the yearly roadmap to these streams. Each roadmap goal is written as a goal tree: the goal, the outcome it requires, the department responsible and the stream that produces it, with the measure that would show progress. A coverage check then lists every yearly commitment beside the stream that serves it, so a commitment with no stream behind it shows up as a gap before the year starts rather than at the end of it.

The map also records how departments depend on each other. In our case R&D is the foundation layer (skills and tooling), Engineering is the production layer, and Content and Business Development amplify what Engineering produces: they need something built before they have something to write about or sell. Recording that dependency explains why an allocation decision in one department changes what the others can do.

Each department receives a share of the week. The strategy agent proposes the shares from the goals and current progress; Richard approves them. The share is the unit the scheduler later turns into blocks on a calendar.

The weekly log and how a department comes online

The weekly tactical cadence (the strategy agent's weekly planning session) writes one entry per week to an append-only weekly log. The entry has a section for each department with three fields: focus (what the department should accomplish), deliverables (specific outputs) and allocation (its share of the week), plus a link to the strategy document of the project it serves. A department with nothing assigned is marked dormant, which is a recorded decision rather than a silence.

When priorities change mid-week, for example because an escalation was accepted, the change is appended to the same week as an update block. The block states what changed, why, each affected department's new priority, what was deprioritised, the implication for other departments and when the change will be re-evaluated. Entries are never rewritten, so the log reads as a history of direction.

Any session that does a department's work starts with the same establishment routine:

1. Read the current week's entry in the weekly log.
2. Check for update blocks added since.
3. Note the department's allocation.
4. Follow the link to the project's strategy document for detail.
5. Query the task tracker for ready work carrying the department's label.
6. Start on the highest-priority item.

The routine means a fresh agent session, with no memory of last week, picks up current direction before it touches any work.

Status flowing back up

Departments report upward through the operational state document: current focus, blockers, health per stream, and a dated activity log of what shipped. Each report is written at the altitude the strategy agent consumes (department health, the headline blocker, results worth recognising) and never copies task-level detail, which the task tracker already holds and which goes stale as soon as work moves.

What exists today

The department structure, activity map, weekly log and establishment routine are in use, with one person filling every department role and Claude Code and Codex sessions doing work inside a department's scope. Dedicated agents per department are the intended next step and are not yet in place; some independent agents already work on specific jobs and report through the file-based agent inbox. How the whole company structure is meant to evolve as agents take on roles is set out in the agentic organisation.