An ecology of mind and machine

"An ecology of mind and machine" is the line under the AR Labs name on the front page of this site, and it states how we think about AI. We treat a person and the AI agents they work with (language models run in a loop with tools, so they can act, not only reply) as parts of one system, together with the documents, task graphs and records they share. We design that system the way an ecologist looks at a habitat: each part has its own strengths and limits, the parts depend on each other, and an action is only good if its consequences are acceptable for the whole.

The phrase echoes Gregory Bateson's Steps to an Ecology of Mind, the book our founding framework, cognitive ecology, draws on most. That page explains the framework from the cognitive-science side. This page explains how we carry it into engineering.

Two minds with matching shapes

The framework describes a human mind as a small, slow conscious attention in front of a vast, fast unconscious process that completes patterns. A large language model has a similar shape: a limited context window in front of a huge store of learned associations, which completes patterns from partial cues and explains itself after the fact.

We draw two consequences. First, the same design rules apply to both. A person's working memory holds only a few items, and a model's attention degrades when its context is crowded, so both need information presented in small, well-chosen pieces, general first and detail on demand. Second, they make good partners. The framework's central idea is that conscious attention works best by framing the work of the unconscious: setting the goal and the boundaries, then letting the faster process fill them in. In a person and model partnership, the person does the framing and the model does the associative work.

flowchart LR
  P["Person: intent, framing, judgement"] -->|"well-formed outcome"| M["AI agents: pattern completion, execution"]
  M -->|"results and evidence"| P
  P --- R["Shared records: goals, task graph, documents"]
  M --- R
  R --> E{"Ecology check: do consequences fit the wider system?"}

The ecology test

"Ecology" in the framework names a specific test: whether what someone intends lines up with the long-term consequences for themselves and the systems around them. Bateson warned that conscious purpose, followed alone, optimises one variable and stays blind to the system that absorbs the cost, and that technology amplifies conscious purpose without widening its view.

We treat AI as that kind of amplifier, and so build the ecology test into our work at several scales. At the scale of a single task, an agent's result is checked against the outcome it was given and against evidence, not against its own report of success. At the scale of a project, we ask whether speeding something up costs learning or quality elsewhere, the argument of the acceleration paradox. At the scale of society, one of our long-term aims is reducing the damaging effects of technology.

One skeleton for people, documents and agents

We use one organising form throughout our system: logical levels, the idea, running from Bertrand Russell through Bateson, that each level of a hierarchy frames and governs the level below. The same skeleton organises Cognitive Ecology, the book manuscript by Richard M. Thompson, who founded AR Labs, written between 2006 and 2010 and now being restructured. It also organises our goal trees, our project documents and our agent instructions.

• Documents. Every project loads in a fixed order: what success looks like, then the current approach, then active work, then individual tasks. People and agents read the same files in the same order; see progressive disclosure.
• Goals. The book's well-formed outcome, a goal stated so precisely that whoever receives it need not guess, stacks into hierarchies: the intention behind an outcome is the outcome one level up, and its first step is the outcome one level down. We are developing it into a contract for handing a goal from a person to an agent.
• Agent instructions. Logical levels applies the same nesting to what an autonomous agent is told, so that a rule at one level does not silently contradict the frame above it.

A system that develops itself

The ecology includes the system's own growth. The agents in VOX, the lab's working system, help build that system: a person designs its structure and methods, and the agents carry out and improve the work within them. Our documents and history form a second brain, plain-text files about our goals and work that people read and agents load each session. When an agent makes the same mistake twice, the fix goes into the environment as a check rather than into a person's memory. Each of these is a feedback loop, the kind of self-correcting circuit Bateson placed at the centre of his account of mind.

Why it matters for the work

Our reading of the framework is hopeful. In a healthy person, conscious and unconscious processes are partners, each doing what the other cannot. Our bet is that people and AI can form the same kind of partnership, on two conditions: the system respects the limits of both minds, and the person keeps hold of framing and ecology instead of handing over judgement. The VOX interface, our context engineering practice and our agent governance are each one part of building that partnership.