Cognitive Movements, Roles, and Carriers

1. Movement is primary

AIOS begins with the cognitive transformation the work requires, not with a permanent agent personality.

A person may need to:

• orient to a situation;
• discover missing evidence;
• decide among consequential alternatives;
• design a plan;
• compose an artifact;
• revise existing work;
• test integrity or evidence;
• integrate what changed;
• reflect on purpose, drift, or completion.

These are judgment forms. They can be carried by a conversational turn, a stable seat, a workflow stage, a temporary role, a delegated branch, or a composed sequence.

This prevents the architecture from assuming that every recurring cognitive act requires a new permanent agent.

It also prevents model selection from occurring at the scale of an entire person or project. The same domain may use a private local model for retrieval and routine revision, a specialized model for technical analysis, and a frontier model for an unusually novel synthesis. The thing being routed is the bounded movement with its evidence, risk, and expected product.

2. Nine judgment forms

The nine forms are a conceptual vocabulary, not a fixed runtime roster.

3. The distinction among form, seat, role, workflow, agent, and command

These terms answer different questions.

Conflating these produces either agent sprawl or rigid orchestration. Keeping them distinct lets the same cognitive form appear in many domains while the same role participates in many workflows.

This distinction also separates the system's operational substrate from intelligence moving through it. A command can admit a capability, a resolver can bind an identity, and an operation can write an exact permitted change without performing semantic judgment. Conversely, a model can make a meaningful comparison or recommendation without acquiring permission to execute it. The architecture needs both layers, but neither layer should be mistaken for the whole intelligence system.

4. A seat is the unit of model judgment

A seat is the complete executable package for one kind of semantic judgment.

It includes:

1. stable identity and responsibility;
2. system prompt in the language of the work;
3. declared context composition;
4. exact source producers for each context region;
5. one normal response grammar;
6. parser and type agreement;
7. a declared consumer or file effect;
8. bounded failure behavior;
9. verification through fixtures, assembled-context evidence, and a reachable product path.

The seat is deliberately narrow:

• a framer frames;
• a writer writes;
• a selector chooses from supplied possibilities;
• an editor changes exact language;
• a tester reads independently;
• a landing seat states what happened and what it bears on.

A seat that writes, routes, audits, plans, and evaluates simultaneously spends its attention on system administration and is difficult to verify.

4.1 The reasoning-turn envelope

A complete reasoning turn has three logical functions even when they are implemented within one invocation or distributed across several carriers:

Attention formation is itself cognitive work. It determines what becomes salient, which relationships become visible, what remains intentionally absent, and what question the judgment is actually answering. It does not need to be a separate model call: exact source resolution, revision checking, and contract assembly may be mechanical, while purpose, relevance, standing, or semantic altitude may require judgment.

The middle function owns one bounded situated judgment. “One” names coherent responsibility, not a claim that the model performs only one mental operation. A Compose movement may compare, draft, and resolve local tension internally while remaining responsible for one declared semantic product.

The landing function must name a consumer. That consumer may be the person, a later seat, an artifact, a workflow stage, a plan, or an integration boundary. It can expose immediate choices while already-authorized deeper work continues asynchronously against its original object. It can also return a bounded unknown or no further recommendation. The absence of a fitting next action should remain visible as an empty action field rather than being filled with invented activity.

5. Mind-language, not container-language

The prompt should name the meaning of the judgment rather than narrating implementation machinery.

For example:

• the architecture may call something a response parser;
• the prompt asks the seat to return one clear decision block;
• the runtime may store a trace field;
• the prompt asks for a concise account of the movement;
• the host may inject a document-map region;
• the prompt describes it as the accepted structure of the adjacent document.

The output vocabulary required for mechanics does not need to become the model's complete cognitive vocabulary. Mechanical identifiers remain outside the model-facing field unless exact reference is necessary.

6. Carriers of intelligence

AIOS carries judgment across boundaries through several related forms.

6.1 System prompts

Stable responsibility, reading discipline, boundaries, and output contract for a seat.

6.2 Context capsules

Freshly composed ground for one invocation: purpose, object, memory, evidence, role, workflow stage, artifact state, and current input.

6.3 Natural-language directives

Judgment handed from one movement to another in the form:

intent
  → few true invariants
    → invitation to solve the actual case

The directive keeps substantive judgment whole while exact identities and effect boundaries remain structured.

6.4 Menus and controlled selections

Known options from which a model may select semantic orientation or a person may select action. The model chooses meaning; deterministic machinery resolves the chosen identity.

6.5 Typed output contracts

Declared forms that convert model results into bounded consequences: natural language plus trailers, delimited blocks, strict YAML, or another seat-specific grammar.

6.6 Files and companion metadata records

Durable carriers that preserve artifact language, standing, structure, decisions, relations, lineage, and reconstruction.

6.7 Receipts and returns

Evidence of what was authorized, attempted, applied, produced, accepted, or left unresolved.

7. Stable core, contextual expertise

The system avoids rewriting every prompt for every domain.

The convergence is the point: specialization lives in composed ground around a coherent responsibility, not in a growing roster of hardcoded agents. The same seat can serve different domains because role, method, evidence, and work state change with the commission.

A research workflow and a strategy workflow can use related reasoning forms while receiving different expertise, evidence standards, terminology, and expected outputs.

This is a key generality mechanism: specialization is injected as natural-language ground rather than compiled into an ever-growing set of hardcoded agents.

8. Recurring executable forms

The Fractal Seed appears in several route families.

These forms share a grammar without becoming one monolithic pipeline.

9. Explore as temporary specialization

Research and context work illustrate why a permanent agent taxonomy is unnecessary.

An Explore directive defines:

• the exact unknown, alternative, context gap, source need, or evidence question;
• the work or decision that the answer can change;
• granted files and authorized wider search;
• settled facts and constraints;
• inquiry method and evidence requirements;
• useful return shape;
• boundaries and stop condition.

A temporary research branch may then:

read the inquiry
  → perform bounded investigation
    → return findings, evidence, contradictions, limits, and delta

The role is configured by the charge. It does not need to exist as a persistent global research agent.

10. Output forms preserve different kinds of result

Three broad product forms must remain distinct.

Language

The reasoning, interpretation, or artifact prose the person reads.

Selections

Declared context or reasoning choices selected from supplied possibilities and resolved mechanically.

Proposals

Bounded possibilities for further work. A proposal becomes an operation only after the appropriate person take-up or prior grant.

A proposal set may truthfully be empty. Empty is a semantic result when no supplied action fits, no consequence is warranted, or the movement should stop. The carrier must not fabricate a menu item merely to satisfy an interface shape.

These output forms are not mapped one-to-one onto Why, How, and What. Any movement may produce language, selections, or proposals.

11. Failure behavior is part of cognition architecture

Malformed output must not cause the system to invent meaning.

Different results have different integrity requirements:

• a valid public response may survive a malformed optional trace;
• an edit batch may need atomic rejection if one locator is stale;
• a partial research return may remain useful if its limitation is explicit;
• a recommendation list may truthfully be empty;
• a failed projection must not be shown as completed work;
• an integration orientation must not invent its own destination.

The failure contract protects both epistemic standing and mechanical effects.

12. When to create a new seat

A new seat is justified when:

• the system needs a genuinely different coherent judgment;
• the context composition differs materially;
• the response contract differs materially;
• independent perspective is required;
• its product has a named consumer that cannot reliably receive the result from an existing seat;
• combining responsibilities predictably degrades quality or verification.

A new seat is not justified merely because:

• a domain has a new label;
• a mechanical branch exists;
• a workflow adds another stage;
• a temporary exploration can be defined by a directive;
• the current prompt could be made clearer.

13. Research implications

This architecture creates several research questions:

• Does one-seat/one-judgment separation improve reliability relative to multifunction agent prompts?
• Which judgment forms benefit from independent landing or testing seats?
• How much domain expertise can be injected contextually before a stable seat should specialize?
• Can directive-based temporary roles match persistent agent personas while reducing orchestration complexity?
• Which output-contract strictness levels best preserve semantic richness without increasing application error?
• How should failure behavior be evaluated when a response contains both useful language and malformed control data?
• Under matched resource budgets, when does explicit attention formation and landing improve accepted outcomes relative to one extended reasoning trace, and when does the additional handoff burden make them worse?

The relevant comparisons must separate cognitive specialization from agent multiplication. A useful result from distinct localization, drafting, and validation stages does not prove that each stage needs a permanent persona. Likewise, several agents using the same evidence and assumptions do not constitute independent review merely because they have different names.

14. Relationship to the whole and research

Cognitive movements operationalize the Fractal Seed without reducing it to a fixed number of calls. Each movement receives a projection of the semantic situation, becomes the receiving unit for context composition, and returns through the authority and integration path developed in Guidance Without Cages. The later workflow chapters show how several movements can cooperate without turning the workflow into the owner of their judgment.

Recent research supports purpose-specific staging while warning against universal agent taxonomies. Agentless showed strong results from explicit localization, repair, and validation phases in a bounded software task. A 2026 study of multi-agent coordination found that coordination gains depend heavily on task structure and can reverse on sequential work. These findings are compatible with the AIOS decision to distinguish transformations by their context and output needs while using the minimum coherent number of models, seats, or branches.

The AI-native architecture research brief contains the evidence and counterevidence in depth. The public claim remains architectural and testable: materially different cognitive movements should be separated when their information needs, contamination risks, output obligations, authority, or validation regimes differ—not merely because different role labels are available.