Cognitive Movements, Roles, and Carriers

1. Movement comes before agent

AIOS begins with the transformation the work requires. It does not begin by assigning every recurring activity to a permanent agent personality.

A person may need to orient to an unfamiliar situation, discover missing evidence, decide among consequential alternatives, design a plan, compose an artifact, revise existing work, test integrity, integrate a result, or reconsider the purpose of the entire effort. Each demand directs attention differently and produces a different kind of semantic result.

AIOS calls these demands cognitive movements. A movement can be carried by one conversational turn, a stable semantic seat, a workflow stage, a temporary role, an asynchronous branch, or a composed sequence. The carrier serves the movement; the movement does not exist to justify the carrier.

This principle prevents agent sprawl. The architecture adds a new permanent seat only when the system genuinely needs a distinct and repeatable form of judgment with its own context, responsibility, and result.

2. Nine judgment forms

The current architecture distinguishes nine recurring transformations.

These forms are the present illustrative vocabulary for cognitive responsibility. They show how AIOS differentiates the judgments required by real work; they are not a closed taxonomy or a fixed runtime roster. The categories can continue to evolve as the architecture develops. One workflow may use several forms, and one form may appear across many domains.

3. Form, seat, role, workflow, branch, command, and operation

These concepts describe different parts of the architecture.

A form identifies the cognition. A seat packages one model judgment. A role supplies contextual expertise. A workflow relates several stages over time. A branch receives a commission and return contract. A command represents person take-up. An operation changes the world mechanically.

Conflating them produces two opposite failures. Everything becomes a generalized agent with expanding responsibilities, or everything becomes a rigid pipeline whose semantic choices were predetermined by software. Distinction lets the system combine open-ended judgment with exact operation.

4. The seat is the unit of model judgment

A seat is the executable package for one coherent semantic responsibility. It contains:

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

Narrow responsibility protects judgment quality. A framer forms attention. A writer composes. A selector evaluates supplied possibilities. An editor changes exact language. A tester reads independently. A landing seat identifies what happened and what it bears on.

A seat that writes, routes, plans, audits, evaluates, and integrates in one invocation spends its attention managing the system instead of performing the judgment. Separation keeps the object's meaning in the foreground and makes the result easier to verify.

5. Mind-language keeps the judgment intelligible

Model-facing language names the meaning of the task rather than narrating the implementation container.

The runtime may call a structure a response parser; the prompt asks for one clear decision block. The system may store a trace field; the prompt asks for a concise account of what the movement established. A host may inject a document map; the prompt presents it as the accepted structure of the surrounding publication.

Mechanical identifiers enter the model's context when exact reference is necessary. They do not replace the vocabulary through which an intelligent participant understands the object, purpose, evidence, and transformation.

Mind-language matters because container-language transfers infrastructure burdens into cognition. A model asked to reason as a queue processor, schema adapter, router, and evaluator at once has less attention available for the semantic work the person actually needs.

6. Carriers move intelligence across boundaries

Each carrier preserves the one thing that must survive a particular boundary crossing. Stable responsibility must cross from architecture into an invocation; fresh situation must cross from durable ground into present attention; substantive judgment must cross between movements; exact identity must cross from language into operation; and accepted consequence must cross back into durable state. AIOS therefore uses different carriers rather than asking one prompt or schema to carry every responsibility.

System prompts

A system prompt establishes the seat's stable responsibility, reading discipline, boundaries, and response contract.

Context capsules

A capsule composes fresh ground for one invocation: intent, active object, memory, evidence, role, workflow stage, artifact state, current input, authority, and expected product.

Natural-language directives

A directive hands judgment from one movement to another as:

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

The directive preserves substantive judgment as a whole while exact identities and effect boundaries remain structured.

For example, a review movement can hand its judgment to an editor without reducing the finding to a flag:

Intent: restore the chapter's whole-system argument after the opening was narrowed.
Invariants: preserve the established thesis; change only the affected framing;
            keep evidence and accepted architecture distinct.
Invitation: read the actual chapter, determine the smallest coherent revision,
            and explain what the change repairs.

The directive carries why the revision matters and what must remain true, while leaving the editor responsible for solving the actual language and structure. File identity, edit permission, and effect validation remain in their exact carriers.

Menus and controlled selections

A menu presents meaningful local possibilities. The model judges their relevance; the person sees and redirects the same field; deterministic machinery resolves the selected identity. Consequence is bounded without reducing interpretation to a keyword.

Typed output contracts

A seat returns a declared form suited to its consumer: natural language, a delimited decision block, structured YAML, selections, proposals, or another bounded grammar. The contract protects the bridge from semantic product to exact effect.

Files and companion records

Artifacts preserve exact language and structure. Companion records preserve identity, standing, purpose, relationship, decisions, lineage, and reconstructable state.

Receipts and returns

Receipts record what was authorized, attempted, validated, changed, or refused. Branch returns preserve product, provenance, limits, unresolved conditions, and the delta proposed to the parent situation.

7. Stable cognition, contextual expertise

AIOS specializes judgment by composing expertise around a stable responsibility.

A research workflow and a strategy workflow can use the same judgment form while receiving different terminology, evidence standards, practices, artifact states, and expected products. The seat's core responsibility remains stable; the situation supplies specialization.

This is a major source of generality. Expertise can live in inspectable natural-language ground and reusable workflows instead of an ever-growing inventory of hardcoded agent personalities.

8. Recurring route families

The Fractal Seed appears through different combinations of orientation, development, landing, and reintegration.

The shared grammar enables coordination. The routes remain different because the object, altitude, quality standard, authority, and semantic product differ.

9. Exploration creates temporary specialization

An Explore directive defines a bounded inquiry rather than invoking a permanent research agent. It specifies:

• the exact unknown, contradiction, alternative, source need, or context gap;
• the work or decision the answer can change;
• granted files and any authorized wider search;
• settled facts and constraints;
• inquiry method and evidence requirements;
• the useful return shape;
• authority, boundaries, and stopping conditions.

The temporary branch then reads the inquiry, performs the investigation, and returns findings, evidence, contradictions, limitations, and proposed delta. Its expertise comes from the commission and ground. Its authority ends with the return contract.

This makes specialized work composable without constructing a permanent global taxonomy of agents.

10. Product forms preserve different consequences

Three broad output forms remain distinct.

Language carries reasoning, interpretation, recommendations, or artifact prose for a person or another movement to read.

Selections identify one or more supplied possibilities. The selection expresses semantic judgment; deterministic machinery resolves the referenced identity and available effect.

Proposals describe bounded future work, changes, or consequences. A proposal remains inert until a person or prior grant authorizes its operation.

Any cognitive movement can produce language, selections, or proposals. The product form does not map mechanically onto Why, How, or What.

11. Failure behavior belongs to the architecture

A malformed control structure must never cause the system to invent meaning or consequence.

Different products require different failure contracts. A useful public explanation can survive the loss of optional trace metadata. An edit batch may require atomic rejection if one locator is stale. A research return can remain valuable when it states its incompleteness. A recommendation list may truthfully be empty. A failed projection cannot appear as completed work. An integration movement cannot invent a destination absent from its commission.

The contract protects both standing and effect. It preserves useful semantic material where safe, refuses ambiguous mechanical action, and makes the failure visible to the next responsible movement.

12. When a new seat is warranted

A new seat is justified when the system needs a genuinely different coherent judgment, a materially different context composition, a materially different response contract, or an independent perspective whose combination with another responsibility predictably degrades quality.

A new domain label, a mechanical branch, an additional workflow stage, or a temporary exploration does not by itself justify a new seat. Those differences can often be expressed through role, context, directive, workflow, or operation while the cognitive responsibility remains stable.

This rule keeps the architecture precise without multiplying permanent agents around every noun in the product.

13. Research implications

The movement architecture creates a concrete evaluation program:

• Does one-seat/one-judgment separation improve reliability over multifunction agent prompts?
• Which forms benefit most from an independent landing, integration, or testing seat?
• How much expertise can enter through context before stable responsibility itself must specialize?
• Can directive-based temporary roles match persistent personas while reducing orchestration complexity?
• Which output contracts preserve semantic richness while maintaining exact operational safety?
• How should mixed returns be evaluated when useful language accompanies malformed control data?
• Does explicit separation among recommendation, selection, authorization, and operation improve human understanding of consequence?

AIOS organizes cognitive architecture around transformations of meaning. Agents, workflows, prompts, and operations become more coherent when each is derived from the movement it exists to carry.