Cognitive ecology for people and AI agents

Cognitive ecology is the framework AR Labs is built on. It holds that the mind has two very different parts, a small, slow, conscious attention and a vast, fast, unconscious process underneath it, and that good results come from using the first to frame the work of the second. "Ecology" names the test that every action has to pass: whether what someone intends lines up with the long-term consequences for themselves and the systems around them. We apply it to working with AI agents, because a language model has a similar shape: a small context in front of a vast store of learned patterns.

The framework comes from Cognitive Ecology, a book manuscript that Richard M. Thompson, who founded AR Labs, wrote with a collaborator between 2006 and 2010 while working as a cognitive scientist. It was never published and is now being restructured. Its ideas turned out to describe working with AI language models closely, and they inform most of what we build.

Two minds in one skull

The book starts from findings in cognitive science.

Conscious attention is small. Working memory, the part of the mind that holds what is being thought about right now, carries only a few items at once and loses them quickly without rehearsal. The book draws on the working-memory research of George Miller and Alan Baddeley, and uses the image of a flashlight that lights a small patch of a very large room.

Most thinking is unconscious. The bulk of perception, memory and skilled action runs outside awareness. Experts are people who have automated huge repertoires of responses through years of deliberate practice, the research of Anders Ericsson. Much of that skill is invisible even to the expert. Research on self-report going back to Richard Nisbett and Timothy Wilson's 1977 paper "Telling more than we can know" shows that people asked to explain a choice often give a fluent account that has little to do with what drove it.

The brain completes patterns. Drawing on the connectionist models of James McClelland and David Rumelhart, the book treats the mind as a system that recognises and fills in patterns from partial cues, rather than one that applies explicit rules. Our maps of the world are built from experience, and the map is not the territory.

Why conscious attention lacks ecology

The book's central argument comes from the anthropologist and systems thinker Gregory Bateson, especially the essays on conscious purpose collected in Steps to an Ecology of Mind (1972). Bateson argued that conscious purpose pulls short, goal-directed sequences out of a mind and a world that are made of loops. Followed alone, it optimises one variable and stays blind to the system that absorbs the cost. In "Effects of Conscious Purpose on Human Adaptation" (1968) he put it bluntly: "Lack of systemic wisdom is always punished."

The book turns this into a practical claim: conscious attention, by design, favours single descriptions, immediate results, narrow focus and the deletion of whatever seems irrelevant. Those traits make concentration possible, and they also hide consequences. Bateson worried that technology amplifies conscious purpose without widening its view, so that a narrow beam of attention gains an engine. The essay series behind the acceleration paradox applies this argument to AI.

The figure draws Bateson's point twice. On the left, purpose cuts a straight line across a loop and optimises the one variable it can see, while the cost lands on a part of the loop it cannot. On the right, an engine drives the same line harder: the beam is stronger, the view no wider, and the cost larger.

Framing: how attention directs the unconscious

If attention is small but the unconscious is large, what matters most is how attention is used. The book calls this framing: setting the boundaries and parameters within which the rest of the mind acts creatively and adaptively. A frame is a definition of context that constrains which behaviours are acceptable. Change the context and you change the behaviour available.

Several tools follow, most of them adapted from neuro-linguistic programming, the practice-based school of communication and change developed by Richard Bandler and John Grinder in the 1970s. Present state and desired state contrasts where things are with a clear outcome, so that unconscious problem-solving has something to work toward. Perceptual positions, set out by Judith DeLozier and John Grinder, view a situation from one's own eyes, from another person's, and from an outside observer. Outcome, intention, consequence checks that the likely results of an outcome match the intention behind it. Back-burner processing elaborates a problem fully, then deliberately shifts attention away and lets the answer arrive later.

The well-formed outcome

The book's culminating tool is the well-formed outcome, another idea from neuro-linguistic programming: a goal stated so precisely that the part of the mind receiving it does not have to guess. The book, in its restructured form, assembles it from eight conditions, each a framing of one aspect of the goal.

Condition	What it frames
Contextualised	Where, when and with whom the outcome holds
Stated positively	What to move toward, not away from
Sensory evidence	What will be seen, heard and felt when it is reached
Resources	The internal states and external means it needs
Within your control	It can be started and kept going without others changing
Multiple pathways	More than one route, so it widens choice
Ecological	Its consequences are acceptable to the self and the wider system
First step	It is achievable, with a concrete action available now

A well-formed outcome can also be stacked. The intention behind an outcome is itself an outcome one level up, and the first step is an outcome one level down, so the pattern builds whole goal hierarchies.

Process over content

The book's methods were deliberately content-free: they describe how a mind works, not what any particular mind is thinking about. A pattern defined over process does not depend on what it runs on. That is why we find the framework applies to language models. A large language model is also a fast, associative engine that completes patterns, and when asked to explain itself it produces a plausible account generated after the fact. When a person works with an AI agent, the person supplies what the book assigns to conscious attention: intent, framing, judgement and the question worth asking. The model supplies the associative horsepower.

flowchart TD
  H["Person: limited attention, intent, judgement"] -->|"frames the work: well-formed outcome"| M["Model: fast associative pattern completion"]
  M -->|"results to check against the outcome"| H
  H --> E{"Ecology check: do consequences fit the wider system?"}
  E -->|"yes"| A["Act"]
  E -->|"no"| H

How we apply it

• Context engineering is framing for a model: choosing what enters its limited context so that its pattern completion works on the right material.
• Progressive disclosure, loading the most general and constraining information first and detail only when needed, organises our documents for both people and agents.
• Cognitive chunking and the VOX interface design respect the small capacity of human working memory.
• Logical levels extends Bateson's thinking into the design of agent instructions.
• The well-formed outcome as a handoff. The book's restructure is developing the well-formed outcome into a contract for handing a goal from a person to an AI agent: specify the outcome well enough, and the agent does not have to fill gaps with something plausible and wrong.

Our broader reading of the framework is hopeful. In a healthy person, the conscious and unconscious mind are partners, each doing what the other cannot. A person and an AI system can form the same kind of partnership, provided the person keeps hold of framing and ecology instead of handing over judgement.