Logical levels and the theory of types applied to AI agent goals

What did Gregory Bateson mean by logical levels, how does the idea draw on Russell and Whitehead's theory of types, and what does it imply for nesting an AI agent's goals, approaches and tasks? AR Labs researched those questions in November 2026; the findings sit behind our concept page on logical levels in agent design. The report works from Bateson's essays and related systems theory (McCulloch, Koestler, Hofstadter, Ashby, Beer, Simon, Watzlawick), and labels each claim as Bateson's own, a secondary interpretation, or our inference.

The design in question is a proposed nesting of frames in the marketing manager agent of our multi-agent system, which runs on the Hermes agent harness (a harness is everything around the model that shapes what an agent does: the agent program, its instructions, skills, hooks, tools and checks): company context, marketing, marketing goals, approaches, then campaigns, routines and tasks. The design questions in section 3 are inputs to a design decision, not decisions.

Three kinds of claim are kept apart throughout. Bateson marks what Bateson wrote. Secondary marks other authors' interpretations. Inference marks our reasoning about the design. Where we are working from memory of a text rather than a source re-read for this note, the entry says so; section 5 lists what was and was not verified.

1. Bateson's own account

1.1 The debt to Russell and Whitehead

Bateson took the idea from Bertrand Russell and Alfred North Whitehead's theory of logical types, set out in Principia Mathematica (1910–13). The theory was built to block paradoxes such as Russell's "the class of all classes that are not members of themselves". Its rule is that a class is of a higher logical type than its members, so a class cannot be a member of itself and a member cannot be the class. A statement about statements sits one type above the statements it is about.

Bateson. In "The Logical Categories of Learning and Communication" (written 1964, reprinted in Steps to an Ecology of Mind, 1972, pp. 279–308) he states the rule in these terms: a class cannot be a member of itself, and the name of a class is of a different logical type from the names of its members. He adds the companion rule he repeats across his work, that "the name is not the thing named" and the map is not the territory.

Bateson. He does not use the theory as formal logic. He says plainly that he uses it as a model or heuristic for the natural history of communication, and that real organisms break it all the time. Play, humour, fantasy, art and schizophrenia all depend on, or suffer from, crossings between levels. In Mind and Nature (1979, chapter on "Every schoolboy knows" and the section "Logic is a poor model of cause and effect") he goes further: logic is timeless, but causal circuits run in time, so a paradox that would be a contradiction in logic shows up in a real system as an oscillation. His example is the electric buzzer, whose circuit says "if the contact is made, the contact is broken".

1.2 Levels of learning

Bateson. The 1964 essay builds a hierarchy in which each level is change in the process of the level below. Paraphrasing his definitions closely:

• Learning 0 is a specific response that is not corrected, whether right or wrong. Receiving a signal and responding the same way every time is Learning 0.
• Learning I is change in the specificity of response by correcting errors of choice within a fixed set of alternatives. Pavlovian conditioning and habituation are his examples.
• Learning II is change in the process of Learning I: a corrective change in the set of alternatives from which choice is made, or a change in how the stream of experience is punctuated into contexts. He also calls it deutero-learning, a term from his 1942 paper "Social Planning and the Concept of Deutero-Learning". This is learning to learn, and its products are character, habits of perception, and premises that are largely self-validating and unconscious.
• Learning III is change in the process of Learning II: a corrective change in the system of sets of alternatives. He says it is rare in humans, occurs in some psychotherapy and religious conversion, and can be dangerous.
• Learning IV would be change in Learning III. He says it probably does not occur in any adult living organism, though evolution and ontogeny may combine to produce it.

Bateson. Two features of this hierarchy matter for design. First, the levels are not categories of content. They are orders of change, which Bateson compares to the Newtonian series of position, velocity (change of position) and acceleration (change of velocity). Second, the higher level is the context of the lower one. A Learning I event happens in a context, and Learning II is learning about kinds of context. He introduces "markers of context": signals that tell the organism which set of alternatives it is choosing from. He also says that a change at a higher level is typically driven by contradictions that accumulate at the level below. Learning III arises out of contradictions in Learning II.

1.3 Frames and meta-messages

Bateson. "A Theory of Play and Fantasy" (1955, reprinted in Steps, pp. 177–193) starts from watching monkeys at play at the Fleishhacker Zoo. A playful nip denotes a bite but does not denote what a bite would denote. For that to work, the animals must exchange a meta-message, "this is play", that is about the other messages and tells the receiver how to read them. A meta-message is a message whose subject is other messages. Metacommunication is communication about communication.

Bateson. He calls the thing set up by such a meta-message a psychological frame, and he lists its properties. A frame is exclusive: by including some messages it excludes others. A frame is inclusive in the same act. A frame is related to premises: the frame around a picture tells the viewer not to use the same kind of thinking inside it as on the wallpaper outside it. His gloss of the picture frame is "attend to what is within and do not attend to what is outside". Any message that defines a frame thereby gives the receiver instructions or aids in understanding the messages inside it. Conversely, every metacommunicative message is or defines a frame. He notes that the mathematical analogy of a set boundary is imperfect, because a psychological frame is a message and therefore has a real existence that a set boundary does not.

Bateson. The play frame is itself paradoxical in Russell's sense, because the messages inside it are partly about themselves, like the Epimenides paradox. He argues that this paradox is not a defect but a necessary condition of play, fantasy and the evolution of communication.

1.4 Paradox and the double bind

Bateson. "Toward a Theory of Schizophrenia" (Bateson, Jackson, Haley and Weakland, 1956, reprinted in Steps, pp. 201–227) describes the double bind as a pattern built from level confusion. Its ingredients are a primary negative injunction ("do not do X or I will punish you"); a secondary injunction at a more abstract level that conflicts with the first, often carried non-verbally ("do not see this as punishment"); and a tertiary injunction that forbids the victim to leave the field. The victim cannot satisfy both levels and cannot comment on the contradiction. The key point for design is the third condition. A contradiction between levels becomes pathological when the receiver is not allowed to metacommunicate about it, that is, to step outside and say "these two instructions conflict".

1.5 Mind and Nature (1979)

Bateson. Of his six criteria of mental process, the last two concern levels. Criterion 5 says that the effects of difference are transforms, or coded versions, of events that preceded them. Criterion 6 says that describing and classifying these transformations discloses a hierarchy of logical types immanent in the phenomena. In other words, he treats the levels as found in how a system codes and responds to information, not as a filing scheme imposed from outside.

1.6 A critique of Bateson's use of types

Secondary. Anthony Wilden, System and Structure (1972), and Wilden and Wilson, "The Double Bind: Logic, Magic and Economics" (in Sluzki and Ransom, eds., Double Bind, 1976), argue that Bateson stretched Russell's theory beyond its logical scope and that real systems are better described as hierarchies of constraint. Bateson's own admission that he uses the theory heuristically (1.1) concedes much of this.

1.7 What goes wrong when levels are mixed (Bateson's catalogue)

Bateson. Across these essays the failures fall into a few kinds. Treating the name as the thing named, or the map as the territory. Treating a message as if it were not framed, as when a nip is read as a bite. Applying the rules for changing one level to a different level, for instance trying to correct a premise (Learning II) by the trial-and-error that works on a single response (Learning I). And contradictions between levels with no permitted route to comment on them, which is the double bind.

2. Related systems ideas that bear on the design

Warren McCulloch, heterarchy. "A Heterarchy of Values Determined by the Topology of Nervous Nets", Bulletin of Mathematical Biophysics 7 (1945), pp. 89–93. McCulloch showed that a nervous system can hold circular preferences (A over B, B over C, C over A) that no single ranking can represent. A heterarchy is an arrangement in which no element is permanently on top and authority depends on the situation. Bearing: links between goals, such as drivers, may form a many-to-many network rather than a tree.

Arthur Koestler, holons. The Ghost in the Machine (1967). A holon is something that is at once a whole to its parts and a part to its whole, facing both ways; a holarchy is a hierarchy of holons. Koestler says each holon is governed by "fixed rules", which he calls its canon, while using "flexible strategies" inside them. Bearing: this is close to the proposed rule that an inner frame cannot loosen the limits of its enclosing frame but chooses freely within them.

Douglas Hofstadter, tangled hierarchies. Gödel, Escher, Bach (1979). A tangled hierarchy, or strange loop, is a system in which moving up through levels brings you back to where you started, as when rules get rewritten by the moves they govern. Hofstadter notes that every such tangle rests on an "inviolate level" that the tangle cannot touch, such as the hardware a self-modifying program runs on. Bearing: if the agent can edit its own frames, something must sit outside its reach.

W. Ross Ashby, ultrastability. Design for a Brain (1952; 2nd ed. 1960). An ultrastable system has two feedback loops. A fast first-order loop adjusts behaviour. A slow second-order loop changes the first loop's parameters only when an "essential variable" goes out of bounds. Bearing: this is Bateson's Learning I and II in engineering form, and Chris Argyris and Donald Schön's single-loop and double-loop learning (Organizational Learning, 1978) is the management version.

Stafford Beer, the Viable System Model. Brain of the Firm (1972) and The Heart of Enterprise (1979). Beer's recursion principle says every viable system contains, and is contained in, a viable system with the same structure. Each level manages variety (the number of possible states) by attenuating what flows up and amplifying what flows down, so higher levels work from summaries. Beer also provides an "algedonic" channel, an alarm signal that bypasses the normal levels when something critical happens. Bearing: summaries of enclosing frames, and an escalation route that skips levels.

Herbert Simon, near-decomposability. "The Architecture of Complexity", Proceedings of the American Philosophical Society 106(6), 1962. Complex systems tend to be hierarchies in which interactions inside a subsystem are much stronger than interactions between subsystems, so each subsystem can be treated in detail while its neighbours are treated in aggregate. Bearing: the best theoretical support for dropping a frame's detail when stepping out of it.

Paul Watzlawick and colleagues. Pragmatics of Human Communication (Watzlawick, Beavin and Jackson, 1967) separates the content level of a message from the relationship level, which classifies the content. Change (Watzlawick, Weakland and Fisch, 1974) separates first-order change inside a system from second-order change of the system itself, explicitly using logical types. These are the most direct bridge from Bateson's levels to practical work on change.

3. Application to the agent's design

All of this section is inference, built on the sources above. In the design under discussion, a frame is one item at one level (a marketing goal, an approach) with everything inside it, and the agent's records (evidence, learnings, procedures and decisions) sit in stores outside the nesting of frames.

3.1 Are the nested frames logical levels, or containment?

As currently described, the nesting is mostly containment: a campaign is a part of an approach, which is a part of a marketing goal's pursuit. Part-of is not class-and-member. A campaign is not a member of the class "approach" in Russell's sense. In shape it is a hierarchy of content categories (company, role, goals, methods, actions), not Bateson's hierarchy of orders of change.

The nesting becomes logical levels in Bateson's sense when each frame is a message about how to read and choose what is inside it, rather than a box that holds it. Concretely, a frame qualifies when it states three things about the level below: the set of alternatives the inner level chooses from (Bateson's Learning II), the criteria by which an inner move counts as an error (what Learning I corrects against), and the process by which the inner level itself may be changed. The proposed rule that an inner frame cannot loosen its enclosing frame's limits already points this way; it is Koestler's canon-and-strategy relation. The play-and-fantasy sense of frame fits better than the learning sense: a frame's summary, carried into context, acts as the meta-message "this work is to be read as part of X, under these premises".

A rough mapping to Bateson's learning levels may help. Running a procedure as written is Learning 0. Adjusting a task or a post in light of its results, inside a campaign, is Learning I. Changing which kinds of campaigns an approach uses, or which approaches serve a goal, is Learning II, and the learnings store is where its products would live. Changing Marketing's mandate or identity is Learning III, which Bateson's account suggests should be rare and, for this design, reserved for the agent's human owner.

3.2 Are the stores of a different logical type from the frames?

Yes in one sense, and the stores are also of several types among themselves. The frames are meta-messages: they tell the agent how to read and act. The stored records are messages: evidence reports events, a learning generalises across evidence, a procedure encodes a habit, and a decision records a choice made under some frame's authority. Bateson's principle that meaning depends on context ("without context, words and actions have no meaning at all", from Mind and Nature) supports keeping records outside the nesting, because the same finding can legitimately mean different things to different frames. What his framework adds is that the citation, not the record, carries the framing: a frame that cites a finding should say what the finding counts as for that frame.

The record types are not all one level. Evidence is first-order description, closest to the territory. A learning is a statement about many pieces of evidence and so is one type up. A procedure is a frozen Learning I or II result. A decision is bound to the frame whose authority produced it, so a decision record should name that frame. Keeping these types distinct, and never letting a record change type silently, is the store-side equivalent of keeping frames distinct.

3.3 Is a driver a cross-level relation?

Yes. A driver links a company goal in the Company Context frame to a marketing goal inside Marketing, and it is a claim about the marketing goal made from the frame above it: "achieving this moves that". Level theory treats relations between adjacent levels as necessary, since each level is the context of the one below. Three consequences follow. First, the driver is a hypothesis, a piece of map, and evidence about whether it holds belongs in the stores. Second, by the rule that a class cannot be a member of itself, the marketing goal's own frame should not be the place that validates or rewrites its driver; the enclosing frame should. Third, in Ashby's terms the company goal is the essential variable: if it does not move, the second-order loop fires and the set of marketing goals is reconsidered at the Marketing level, not patched from inside a campaign. McCulloch's heterarchy suggests drivers may be many-to-many, so the design should not assume one marketing goal per company goal.

3.4 Failures from mixing levels to guard against

• A lower frame rewriting a higher one. A task or campaign changing a goal or the mandate. Bateson's account says lower-level contradictions are what prompt higher-level change, so the remedy is that lower frames may raise a contradiction but the change is made by the enclosing frame's process.
• Evidence treated as a goal. A metric that was gathered as evidence becomes the thing pursued, which is the map taken for the territory (and, outside this literature, Goodhart's law).
• A learning treated as evidence, or a procedure treated as a mandate. A record silently changes type and gains authority it was not given.
• Self-validation. A frame defining or grading its own success criteria, which is the class-as-member-of-itself error.
• Lost frame markers. After context is dropped, the agent reads content without the frame that tells it how, like a nip read as a bite. The enclosing-frame summary is the agent's marker of context and must always be loaded.
• Double binds. Two frames issue conflicting instructions at different levels (for example a mandate to be bold and a routine forbidding unapproved posts) and the agent has no permitted route to comment. An escalation route to the owner serves as Beer's algedonic channel and removes the double bind's third condition.
• Wrong rules for the level. Trying to fix a premise by trial and error on tasks, or burning a goal-level review on a task-level error.
• Unbounded meta-levels. Adding frames about frames without end. Bateson finds few real levels above II; Hofstadter's inviolate level argues that the outermost rules, set by the owner and the harness, should be outside the agent's power to edit.

3.5 Does stepping in narrow context and stepping out widen it?

Partly, with one correction. Bateson's picture frame ("attend to what is within and do not attend to what is outside") supports stepping into a frame as narrowing attention. Simon's near-decomposability and Beer's variety attenuation support carrying neighbouring and enclosing frames only in summary. Type theory supports the view that working at a higher level means working with classes and aggregates, not with individual members, so less detail and broader scope.

The correction is that, for Bateson, the inner content has no meaning without its context, and Learning II premises operate exactly by being taken for granted from the enclosing frame. So stepping in can drop the detail of siblings and of the outer frames' contents, but it cannot drop the enclosing frames' premises; those must arrive as the summary. Stepping out widens scope but should not reload inner detail; the outer frame should see inner frames through their reports.

4. Sources

Primary, Bateson:

• Bateson, G. (1972). Steps to an Ecology of Mind. Chandler; reissued University of Chicago Press, 2000 (ISBN 9780226039053). Contains "A Theory of Play and Fantasy" (1955), pp. 177–193; "Toward a Theory of Schizophrenia" (with Jackson, Haley, Weakland, 1956), pp. 201–227; "The Logical Categories of Learning and Communication" (1964), pp. 279–308; "Social Planning and the Concept of Deutero-Learning" (1942).
• German translation of "The Logical Categories of Learning", hosted by vordenker.de: https://www.vordenker.de/vgo/Logische_Kategorien_des_Lernens.pdf
• Bateson, G. (1979). Mind and Nature: A Necessary Unity. Dutton.
• Summary of the four learning levels, P2P Foundation wiki: https://wiki.p2pfoundation.net/Four_Levels_of_Learning

Systems theory:

• McCulloch, W. S. (1945). "A Heterarchy of Values Determined by the Topology of Nervous Nets". Bulletin of Mathematical Biophysics 7, 89–93.
• Koestler, A. (1967). The Ghost in the Machine. Hutchinson.
• Hofstadter, D. (1979). Gödel, Escher, Bach: An Eternal Golden Braid. Basic Books.
• Ashby, W. R. (1952; 2nd ed. 1960). Design for a Brain. Chapman & Hall.
• Beer, S. (1972). Brain of the Firm; (1979). The Heart of Enterprise. Wiley.
• Simon, H. A. (1962). "The Architecture of Complexity". Proceedings of the American Philosophical Society 106(6), 467–482.
• Watzlawick, P., Beavin, J. and Jackson, D. (1967). Pragmatics of Human Communication. Norton.
• Watzlawick, P., Weakland, J. and Fisch, R. (1974). Change: Principles of Problem Formation and Problem Resolution. Norton.
• Argyris, C. and Schön, D. (1978). Organizational Learning: A Theory of Action Perspective. Addison-Wesley.
• Wilden, A. (1972). System and Structure. Tavistock.
• Giunta, I. (2026). Paper on Bateson's learning hierarchy as a multilevel emergent process, Ca' Foscari repository: https://iris.unive.it/handle/10278/5121148 (not read).

5. Confidence and gaps

The Bateson material in section 1 is reported from his texts as we know them; the learning-level definitions match the excerpts found during this research, but we did not have the English full text open, so wording is paraphrase and page numbers come from standard editions. Wilden's critique is summarised from memory and should be checked before being quoted. Section 3 is our inference throughout.