AI strategy agent: planning reviews from the year down to the day

The strategy agent is the AI agent AR Labs uses to plan its work. It runs the lab's recurring planning reviews with Richard M. Thompson, who founded AR Labs, keeps the record of what was decided, and passes the resulting direction to the rest of the work; Richard keeps every strategic decision. An AI agent is a language model run in a loop with tools, so it can act as well as reply. Internally the strategy agent is called the CEO agent, after the seat it fills in the lab's structure. This page describes it as a project: its role, what it reads and writes, how it is kept to strategy, and what runs today. The method it follows is described on strategy agent planning cadences, and the two skills that implement it on strategy agent skills.

Its role

The role is written down as four functions.

• Setting direction. Turn the long-range vision into objectives for the year and the quarter, a theme for the month and priorities for the week, and recommend a change of course when conditions change.
• Allocating attention. Decide which area of work gets the day. While one person directs many agents, the scarcest resource the agent allocates is that person's time, divided between departments: areas of work such as engineering and content, not teams of people.
• Watching progress. Track progress against the objectives, report it at the weekly and monthly reviews, and flag risks and blockers early.
• Coordinating. Run the reviews, settle dependencies between departments, and decide the escalations filed to it.

Its authority is set on a five-level delegation scale. It may settle routine prioritisation itself. On strategy it researches and recommends, and Richard decides.

How it works

The agent is a role loaded on request, not a process that keeps running. A person starts it in a Claude Code session with a slash command. Given a question, it answers through the strategic lens and stops; given no question, it runs whichever review is due. A cadence is one of those recurring reviews, and the cadences form a state machine: yearly, quarterly, monthly, weekly and daily, plus an escalation state that can interrupt at any time. A small state file records when each cadence last ran, so the agent can tell which is due.

flowchart TD
  P["A person"]
  SK["Strategy agent: a skill in a coding session"]
  SF["State file: which review is due"]
  OS["Operational state, written by departments"]
  ES["Escalations"]
  SS["Strategic state, written by the agent"]
  ML["The agent's memory log"]
  DS["Department sessions"]
  P <-->|"one question at each checkpoint"| SK
  SF --> SK
  OS --> SK
  ES --> SK
  SK --> SS
  SK --> ML
  SS --> DS
  DS --> OS
  DS --> ES

Everything it knows comes from files read at the start of a run, in a fixed order: today's date, its role and communication style, the vision, mission, roadmap and current goals, then its state file, the strategic state and the latest entries of its memory log. A session can therefore be cleared or compressed without losing anything the agent needs.

What it reads. The operational state, which department sessions write: each department's focus, its blockers and an activity log of what shipped. Counts and health from the task tracker, never individual tasks. Escalations filed to it through the escalation workflow. Replies from other agents reach it as entries a session has routed into the operational state.

What it writes. The strategic state: the current theme, the week's priorities, key decisions and strategic risks, which the rest of the work reads. The outputs each cadence owns, such as the month's goals or the weekly log. Its state file. And its memory log, a dated record of decisions, lessons and patterns that only the agent writes. Other sessions report up through the operational state and never write into its memory; the reasons for one writer per file are set out in interface contracts between AI agents.

How it reaches other agents. Through the shared mailbox. A strategy session that needs something from the marketing manager agent writes a dated brief into the mailbox, and the reply comes back as a file. The strategy agent has no identity of its own in the lab's shared messaging channel.

How it stays strategic

Three rules keep the agent at the level of direction.

Direction, not execution. When a person names a piece of department work during a review, the agent records it as that department's focus and stops. It does not open the project's files or write into a department's workspace; the work runs in a department session with a fresh context. Execution inside a strategy session fills its context window with project detail the next decision does not need, and it can duplicate a department session already working on the same files.

One question at a time. At each checkpoint in a review (after presenting data, before proposing priorities, before writing outputs) the agent stops and asks one question. Without the pauses an agent works through every phase on its own, and the review becomes a report nobody responded to.

A decision carries its cascade. An approved change of direction is complete only when it names every downstream document that must change, as a checklist the weekly review tracks. A decision recorded only at the strategic layer does not reach the pages, templates and profiles that people outside the lab see.

Design lessons from its failure modes

We keep a named catalogue of the failure modes we have observed in this agent, such as deciding from the most recent document instead of the authoritative one, or grading progress by an easy number that measures something else. The method page lists all seven. Two lessons from the catalogue shape the next version.

First, each failure so far was caught by the person reviewing, who acted as a critic, an authority gate or a reconciliation check. A strategy agent that someone else runs, without that person in the loop, needs each of those as a built component: forced loading of the authoritative documents before a strategic decision, a check that reconciles how a fact is classified in operational and strategic documents, a monitor on dated commitments, a permission model that turns a ruling above the agent's level into a recommendation, and a critic pass before any durable write.

Second, we keep two explanations open. The narrower one is a missing procedural step, which a checklist gate fixes. The wider one is that the instructions have grown more complex than a single context window, with no enforced route through the document hierarchy, can reliably hold, so the agent anchors on whatever is already in its recent context. If the wider explanation holds, the fix is structural rather than procedural. Each new case is recorded against both.

A third lesson came from trying to run the agent without being asked. A skill runs only when someone starts it, so in March 2026 we ran the strategy agent as an always-on process on a server, reaching the person through a messaging app. The runtime woke the agent every half hour and passed each result through a built-in filter that suppresses routine status messages. Reminders about an overdue review are repetitive by nature, so the filter classed almost all of them as routine and nothing reached the person. We stopped the service. Before it runs again it needs scheduled delivery instead of polling, a delivery path tested end to end before any schedule is switched on, and a daily limit on model use.

What exists today

Running: the strategy and escalation skills in Claude Code, a written procedure for every cadence, and state files under version control. The state-machine form has been in use since February 2026, with the weekly, monthly and quarterly reviews in regular use. A weekly memory-consolidation pass, Dream, writes a synthesis of the week's sessions into the agent's memory for the weekly review.

Paused: the always-on process described above.

Designed and not running: a voice interface, a persistent agent process, and an event-driven mode in which the agent reacts to changes in a continuously updated model of the lab's work rather than waiting for a cadence boundary. In that design the cadences remain as review checkpoints, the authority levels still apply, and a person still makes the strategic decisions.

The strategy agent is one of the agents listed on AI agents at AR Labs.