Communicative Agency: Speech Acts Within OODA Control
Dar Aystron Independent Researcher
Abstract
As agents become more capable - richer world models, longer planning horizons, attachment to multiple entities, operation in shared environments - communication becomes structurally necessary, not optional. An agent that plans but cannot share the plan, that detects risk but cannot warn, that commits but cannot coordinate, that learns but cannot teach, hits an architectural ceiling. Communication is not a feature added to agency. It is what enables agency to scale beyond the individual.
This paper develops the role of language and communication within OODA Agency Theory. We argue that communication is not a separate system layered on top of cognition but is integrated into the same control architecture that drives all of the agent’s behavior. Speech acts are actions in the full sense: a :speak may serve as the primary commitment of a cycle’s action bundle or as a subordinate action accompanying a non-communicative primary commitment. Its cognitive grounding is drawn from across the whole cycle and from prior cycles, traceable through the propositional layer introduced in Paper 06. The cognitive work that grounds communication - the ongoing non-verbal work of lifting, evaluating, and reflecting over stream contents - happens at the propositional layer as cognitive flow, and reflection, internal speech, and external speech are three forms through which some of that cognitive flow can surface in linguistic form, distinguished by whether they are substrate verbalization or committed speech actions.
OAT agents model dialogue structure explicitly at the propositional layer, representing causal relationships between speech acts - interruptions, follow-ups, commitments, corrections - as inspectable structure rather than latent pattern. Speech acts are evaluated through the same attachment-weighted mechanism (Paper 07) that governs all action. At the multi-agent scale, speech couples closures in the present; stories couple them across time - extending what coupled closures can do together beyond what any single closure can sustain alone.
1. Introduction
The relationship between communication and cognition in agent architectures has taken different forms over time.
In traditional agent systems, communication occupies a dedicated subsystem. A dialogue manager handles turn-taking, response generation, and conversational state, while a separate cognitive layer handles planning, reasoning, and action execution. The two systems interact through narrow interfaces. Communication is treated as auxiliary to cognition - architecturally separate and functionally secondary.
Contemporary large language model agents appear to have moved to the opposite extreme. In these systems, the control loop is language. Reasoning is chain-of-thought text generation. Planning is produced as natural language. Action is tool-calling described in prose. It looks like the separation between internal cognition and external expression has collapsed - everything operates on the same linguistic surface.
These two positions frame a design space. The legacy separation keeps communication from drawing on the agent’s full cognitive process. The language-centric approach integrates them but may not maintain a clear distinction between private cognitive state and communicative output.
OAT occupies a specific position in this space. Communication is deeply integrated with the cognitive architecture - not separated into a subsystem - but the agent maintains a distinction between internal process and communicable output. The propositional lift (Paper 06) is the mechanism by which selected implementation-layer cognitive contents become explicit propositions in the subject stream. The continuous two-layer activity of the OODA cycle - substrate work at the implementation layer together with the propositions the lift surfaces into the subject stream - is what we call cognitive flow. Cognitive flow runs as the cycle perceives, orients, decides, and acts, and is mostly non-verbal: it surfaces in linguistic form only when verbalization serves the cycle’s work. When cognitive flow surfaces in linguistic form, it does so as one of three categories: reflection (substrate verbalization that enters the stream without being committed as an action), internal speech (a committed :speak action addressed to :Self), and external speech (a committed :speak action delivered to another agent during the Act phase). Internal and external speech share the same Paper 04 commitment structure and differ only in recipient. Reflection differs from both: it arises from the lift at any step boundary rather than being committed by Decide, and it accompanies cognitive flow as substrate-level verbalization rather than as action. Cognitive flow is primary; the three verbalized forms are how some of that flow surfaces in language. The agent has genuine private process alongside genuine communicative capability, and the boundary between them is structural rather than accidental.
This paper develops the consequences of that integration. §2 argues that communication is a structural requirement of advanced agency. §3 situates speech acts within the OODA loop, including their role as primary or subordinate actions in the bundled execution structure of Paper 04. §4 develops cognitive flow as the primary cognitive activity of the loop and positions reflection, internal speech, and external speech as three forms through which it surfaces in language, walking through the pipeline from cognitive flow to committed speech act to rendered utterance. §5 introduces explicit dialogue structure at the propositional layer. §6 connects communicative acts to attachment-based evaluation and develops communication-specific critics over the propositional structure. §7 draws on prior implementation experience with speech-act-integrated agent architectures. §8 extends the analysis to multi-agent systems, arguing that communicative coupling - in real time through speech, across time through story - is the mechanism through which agency persists beyond the individual.
2. Communication as a Structural Requirement
As an agent’s capabilities grow - richer world models, longer planning horizons, attachment to multiple entities, operation in environments shared with other agents - the ability to make internal cognitive state selectively visible becomes not merely useful but necessary. An agent that can plan but cannot coordinate, that can detect risk but cannot warn, that can learn but cannot teach, that can commit but cannot negotiate, encounters limits that are not resource constraints but architectural ceilings. Communication is what enables agency to scale beyond the individual.
The intellectual foundations of OAT’s communication component draw on multiple traditions: speech act theory [1][2] established that utterances are actions; work on shared intentionality and cultural cognition [3][4] showed that human cognitive architecture co-evolved with communicative capacity; work on inner speech [10] established that linguistic forms internalized from social interaction become part of how the mature agent thinks; formal discourse representation [5] and ontological semantics [6] pursued explicit, computable representations of linguistic meaning; computational approaches to dialogue [7] bridged theory and implementation; and classical multi-agent systems research [13][14], particularly the BDI (belief-desire-intention) architecture [15], developed formal inter-agent communication grounded in speech act taxonomies. OAT contributes the integration layer - a control architecture that specifies where communication sits in cognition, how internal state becomes communicable, and why the agent communicates what it does to whom.
3. Speech Acts as OODA Actions
Speech act theory, originating with Austin and developed by Searle, established that utterances are not merely carriers of information but acts in their own right - assertions, commitments, directives, declarations, and expressives - each carrying illocutionary force that creates obligations, changes states, or restructures relationships between speakers [1][2].
Subsequent work in multi-agent systems and computational linguistics extended speech act theory toward agent architectures - notably FIPA-ACL and BDI-based communication protocols. These efforts gave agents structured communicative capabilities, but communication remained a coordination mechanism between agents rather than a surface of each agent’s internal cognitive process. OAT integrates speech acts into the control loop itself.
Within the OODA loop [8], speech acts are actions in the full sense developed in Paper 04: they are constructed as candidates during Orient, committed to during Decide, and executed during Act. Once executed, they are irreversible - a promise made, an assertion stated, a question posed cannot be unspoken. They enter the agent’s operational history and constrain subsequent OODA cycles.
3.1 Speech Acts Fire from Act, Grounded in the Whole Cycle
A speech act is an action. As §3.3 develops in detail, it occupies a position in the action bundle that Paper 02 (Section 7) and Paper 04 establish for any cycle’s commitment structure. Like any action, it is constructed as a candidate during Orient, committed to during Decide, and executed during Act. It does not fire from Observe, and it does not fire from Orient or Decide independently of an action commitment. The communicative output of the agent comes from Act, where it is bound as either the primary commitment of the cycle’s bundle or as a subordinate action accompanying a non-communicative primary commitment.
What distinguishes communication from other kinds of action is not where it fires but what it draws on. The cognitive grounding of any speech act - the contents that explain why the agent is saying this - is drawn from across the whole cycle, and from prior cycles. Perceptual content lifted during Observe (:see, :distance, :approaching), interpretive content lifted during Orient (:risk, :dangerous, :assessment), decision records lifted during Decide (:decision, :selected), and action records from earlier in the cycle or from prior cycles available via internal sensing - all of these are stream propositions that a speech act can reference through its :about role (developed in §5). The agent’s communicative behavior is therefore cognitively grounded in the full propositional context of the cycle that produced it, even though the act itself fires only from Act.
This is the architectural meaning of “communication embedded in cognition” rather than “communication layered on top.” A separate dialogue manager would generate communicative output by consulting a narrow interface to the cognitive system - perhaps a summary of the agent’s current goals, perhaps a recent conversational state. The OAT agent generates communicative output through the same action mechanism that generates all other actions, drawing on the same cognitive flow that the rest of the cycle has been building. The dialogue manager, in OAT terms, does not exist as a separate module. The OODA loop is the dialogue manager.
A separate observation: reflection is not constrained in the same way. Reflection is not an action and does not fire from Act. As §4 develops, reflection is a lifted stream proposition - substrate verbalization that the cycle can produce at any step boundary in any phase, Observe, Orient, Decide, or Act. Where reflection arises within the cycle determines whether it influences the current cycle or only subsequent cycles (§4.3), but its production is not bound to any single phase. The asymmetry between speech acts and reflection is exactly this: speech acts fire from Act because they are actions; reflection may arise at any step boundary because it is not.
3.2 Speech Acts as Commitment-Creating Actions
Paper 04 established that action realization creates irreversible commitments that enter the agent’s operational history. Speech acts exhibit this property directly.
A promise creates a temporal commitment (Paper 02) that persists across OODA cycles and must be perceived in subsequent observations for closure to be maintained. A directive (“Please send me the report by Friday”) creates an expectation that, if adopted by the receiving agent, becomes a commitment in their OODA loop. An assertion (“The system is stable”) commits the speaker to a claim whose falsification carries consequences for trust and credibility.
These are not metaphorical commitments. They are operational commitments in exactly the sense defined in Paper 02: internal states that constrain future behavior across cycles. A promise to call someone tomorrow functions identically to an internally generated plan to call someone tomorrow - both persist as commitments that must be perceived in subsequent Observe phases for the agent to remain operationally closed.
3.3 Speech Acts as Primary or Subordinate Actions
Paper 02 (Section 7) observed that an action may consist of “a coordinated bundle of actuator operations executed as part of a single decision,” constituting one action from the perspective of the control loop because they realize a single commitment and produce a coherent effect. Paper 04 elaborates this bundle structure: each OODA cycle resolves to a single primary action commitment that binds the agent’s next causal stance, and this primary commitment may be accompanied by subordinate or internal actions in support. Subordinate actions occur in parallel or sequence with the primary commitment but do not independently bind the agent or define alternative futures. The structure is neutral with respect to the modality of action - the primary commitment may be physical, computational, mental, or communicative, and the bundle may combine modalities freely.
Speech acts can occupy either position within this bundle, and the architectural symmetry is the point.
Speech act as primary commitment. In some cycles, communication is the agent’s principal causal stance toward the world. The agent’s commitment is to say something - to make the promise, to deliver the warning, to issue the directive, to render the assessment. The speech act is what changes the world; everything else the agent does in that cycle (logging the utterance, recording the addressee, updating its model of the conversation) is subordinate to it. A negotiator entering a critical exchange, a controller issuing a clearance, a clinician delivering a diagnosis - in each case the speech act is the primary commitment, and any accompanying physical or computational activity supports it.
Speech act as subordinate action. In other cycles, the primary commitment is non-communicative, but a speech act accompanies it. The pilot executes the evasive maneuver and simultaneously calls out the threat. The autonomous system invokes a tool and simultaneously announces what it is doing. A surgeon performs the incision and simultaneously narrates the step. The speech act is real - it carries illocutionary force, creates commitments, and is evaluated by the same machinery as any other speech act - but it does not independently bind the agent, does not define an alternative future, and does not constitute the cycle’s primary commitment. The maneuver, the tool call, the incision is what defines the cycle’s next causal stance; the speech act participates in the bundle that realizes that stance as a subordinate member of it.
The asymmetry between these two cases is not in the speech act itself but in what role it plays in the cycle’s commitment structure. The same utterance - “I’m taking the left lane” - may be a primary commitment when issued as a unilateral declaration that constitutes the agent’s stance, or a subordinate annunciation accompanying the primary commitment to actually steer left. The illocutionary content can be identical; the architectural role differs.
Two consequences follow. First, the agent’s communicative capacity is not confined to cycles where it has “decided to talk.” Speech acts can accompany any primary action, which means communication is woven through the agent’s behavior continuously rather than sitting as a separate mode of operation. Second, evaluation must consider both roles. The attachment-weighted evaluation developed in §6 applies to subordinate speech acts as much as to primary ones - the verbal warning shouted during the maneuver may carry as much consequence for trust and coordination as the maneuver itself, and the agent’s selection of whether to say it, what to say, and how is part of how the cycle is composed, not an afterthought to the primary commitment.
4. Cognitive Flow, Reflection, Internal Speech, and External Speech
4.1 Cognitive Flow as the Primary Substrate
Paper 06 introduced two representational layers: the implementation layer, containing runtime control structures, and the propositional layer, containing explicitly lifted cognitive contents available for reasoning and reflection. The propositional lift transforms selected implementation-layer contents into explicit symbolic propositions, which accumulate into the subject stream - the continuously evolving representation of the agent’s cognitive activity. In Newell’s terms [9], the lift is the move from the symbol level to the knowledge level: cognitive contents are made available in a form that can be reasoned about in terms of their content rather than their implementation.
Within this architecture, an OODA cycle is realized as a structured sequence of computational steps organized into four phases corresponding to Observe, Orient, Decide, and Act. Each step performs a specific operation on the agent’s runtime cognitive state. At step boundaries, the propositional lift may produce zero or more propositions that enter the subject stream. The contents of these propositions vary with the kind of step that produced them: a perception step naturally lifts observations, an evaluation step lifts assessments, a decision step lifts a decision record, an action execution step lifts an action record. This is the ordinary operation of the lift as Paper 06 describes it.
The lift is selective - Paper 06 is explicit that most implementation-layer activity remains below it and continues to function purely as control. We refer to the cycle’s combined work as cognitive flow: the continuous two-layer activity of the cycle. At the implementation layer, steps in every phase - Observe, Orient, Decide, and Act - perform substrate work: computation, simulation, tool invocation, data manipulation, whatever the step’s processing requires. At the propositional layer, some of these steps return propositions that the propositional lift adds to the subject stream - observation propositions, candidate and scenario propositions, evaluation propositions, commitment propositions, action records, and reflective propositions about prior stream contents. The two layers run in parallel: substrate work flowing forward as updated implementation state, propositional content accumulating in the subject stream as its residue. Orient is where the bulk of cognitive flow happens, but cognitive flow is not confined to Orient - any step in any phase contributes to it. Cognitive flow is what the loop does across both layers.
Most propositions in cognitive flow are not linguistic. When an agent evaluates a situation, weighs alternatives, recognizes a pattern, or identifies a gap in its knowledge, the lifted propositions that carry this work need not take verbal form. They are symbolic structures in the sense Paper 06 develops, expressing the agent’s observations, evaluations, needs, intentions, commitments, action records, and reflections - and they do their evaluative work through the agent’s WBM and ATTM machinery without requiring any rendering into natural language. A great deal of cognition proceeds this way: silent, propositional, non-verbal.
Against this substrate, reflection, internal speech, and external speech are three categories of cognitive content that can take linguistic form. Reflection arises as substrate activity; internal and external speech are actions in the Paper 04 sense, distinguished from each other only by recipient.
Reflection is lifted stream content that surfaces as verbalized commentary on the cycle’s ongoing work. It is not an action - see the example below. It arises from the lift operating on the cycle’s substrate activity, producing propositional content that may take linguistic form. Reflection can serve as a compressed trace of recent cognition, as a verbalization that makes complex propositional structures easier to hold in working attention and act on, or as running commentary on what the cycle is observing, evaluating, or doing. Reflection enters the subject stream alongside ordinary domain propositions and is causally influential there: it can shape later steps within the same cycle when produced early enough - Orient surfacing it as part of new situational state, Decide selecting action candidates that the reflection helped make salient - or shape subsequent cycles when produced too late for this cycle’s downstream phases.
Internal speech is an action. An internal speech act - a :speak proposition with :Self as recipient - goes through the same candidate-construction, commitment, and execution pipeline as any other action. The agent constructs candidate self-directed utterances during Orient, commits to one via Decide, and executes it during Act. Internal speech creates operational commitments, enters the agent’s action history, and can be the primary or subordinate commitment of the cycle’s action bundle. It is a mental action in Paper 04’s sense - an internal operation whose effect is on the agent’s own mental state. Examples include deliberate self-instruction, rehearsal of a candidate external utterance before committing to external delivery, self-reminders of prior commitments, and extended self-guidance during complex deliberation.
External speech is also an action. A :speak proposition with a non-Self recipient: its :to role identifies the recipient, its :type role classifies the speech act (assertion, warning, promise, directive, question), its :content carries the semantic payload as a propositional structure, and its :utterance is the natural-language rendering delivered to that recipient. External speech enters the Act phase as a primary or subordinate action within the cycle’s bundle (§3.3), uses the standard action mechanism that Paper 02 (Section 7) and Paper 04 establish for any action, and creates the operational commitments described in §3.2. It is architecturally symmetric with internal speech and differs only in the recipient.
The distinction between reflection and internal speech is architectural, not phenomenal. The same utterance - “I have to do this” - can arise as reflection (substrate, a verbalized trace that surfaced during the cycle’s reflective work) or as internal speech (action, a committed self-instruction the agent constructed, decided on, and executed). The difference is whether the cycle’s Decide phase committed to producing it as an action. Reflections just surface; internal speech acts are chosen. Subsequent cycles can distinguish them by the presence or absence of an action record in the stream. Vygotsky’s concept of inner speech [10] was developed at a phenomenological level that did not draw this committed-versus-substrate distinction; we hypothesize that what Vygotsky called inner speech corresponds, in OAT terms, to both reflection and internal speech taken together - reflection for the continuous-accompaniment aspect, internal speech for the active self-regulatory aspect.
4.1.1 From Cognitive Flow to Utterance
Because external speech is an action grounded in prior cognitive flow, its production follows a three-stage pipeline: cognitive flow → commitment → execution. This pipeline mirrors the action lifecycle Paper 04 establishes for all actions: candidates are generated during Orient, one is selected as a commitment during Decide, and the committed action is executed during Act.
Consider an agent that recognizes, in the course of its Orient processing, that it needs a piece of information from the user in order to proceed with its current plan. What happens first is not a speech act. It is a lifted proposition at the propositional layer that represents the cognitive gap:
(:need (:as .n1 :in .S1)
:agent :Self
:content [:information (:about .u1 :from :User)]
:because .o7)
where .o7 is an earlier stream proposition that established why the information is needed, and .u1 is the topic the gap concerns. This is pure cognitive flow. No speech act exists yet; the gap is simply visible in the stream.
Orient then generates candidate actions that could close the gap. For the “ask the user” candidate, Orient constructs a fully formed :speak with both its semantic payload and its surface rendering populated, using the agent’s language capabilities together with its model of the audience, the dialogue state, and contextual factors such as register and politeness:
(:speak (:as .sp1 :in .S1)
:agent :Self
:to :User
:type :question
:about [:set .n1 .o7]
:content (:query
:whom :User
:property :deadline
:subject .u1)
:utterance "What's the deadline for the report?")
Three roles carry distinct architectural work. The :about role points back at the cognitive flow that grounds the speech act - the need .n1 and its basis .o7 - explaining why the agent is asking. The :content role carries what the speech act says, structured as a propositional :query rather than a string. The :utterance role contains the surface rendering of that content as the words the agent would actually say.
Decide evaluates the candidate set against ATTM attachments and WBM-projected trajectories, weighing the cost of interruption, the user’s current load, the relationship’s state, the timeliness of the information, and the reliability of alternatives. The surface form matters to this evaluation, not only the semantic payload: the same :content phrased bluntly, diplomatically, or apologetically projects different trajectories through the user’s likely responses. Decide commits to one candidate.
Act executes the committed :speak - delivering the :utterance to the recipient’s sensor field as a causal event in the world. The action enters the agent’s operational history; the words cannot be unsaid.
The journey from recognizing a need to executing a speech act about it may take a single OODA cycle or many. In the compressed case - a responsive conversational turn, a routine acknowledgment, an urgent warning - recognition, candidate construction, commitment, and execution all happen within one cycle. In the more elaborate case, the need is recognized in one cycle but the speech act is not yet committed: subsequent cycles re-encounter the need through the persisting stream content, Orient may refine candidate :content and :utterance variants across cycles, and Decide may commit to other actions in the meantime - gathering more information, waiting, deferring - before finally committing to the speech act in some later cycle. Throughout, the pipeline runs on cognitive flow regardless of whether any of it also surfaces as reflection or internal speech.
4.1.2 Verbalization under Time Pressure
The three-stage pipeline describes the default case in which cognitive flow has time to run before committing to external speech. At high cycle rates, the pipeline compresses. Under combat tempo [8] or emergency response, cognitive flow runs on pattern recognition with minimal reflective lifting, and internal or external speech, if any, is terse and drawn directly from perceptual content. Both reflection and committed speech acts are largely absent in these modes - not because reflective work has been switched off, but because the cycle rate leaves little room for verbalized stream content of either kind. At low cycle rates, by contrast, reflective lifting is prominent, reflection frequently accompanies cognitive flow as substrate commentary, internal speech becomes feasible as the cycle has time to commit to self-directed speech acts, and external speech can be elaborated across multiple cycles before commitment. These are not different architectures. They are the same loop running at different rates with different lift selectivity.
This applies to both internal and external speech. A :speak - whether directed to :Self or to another agent - can be produced directly from trained pattern recognition: greetings, routine acknowledgments, habitual phrasings, and conventional turn-taking on the external side; routine self-reminders, habitual self-corrections, and automatic self-instructions on the internal side. In both cases the speech act is committed without a deliberative middle stage between perception and utterance. The :about references on a pattern-based :speak are thinner than those on a deliberatively constructed one, but they are still present, and the cognitive grounding critic (§6.3) still distinguishes thin-but-valid grounding from missing grounding.
4.2 Shared Cognitive Flow
Reflection, internal speech, and external speech all draw from the same cognitive flow: the prior propositional content of the cycle and the stream contents accumulated across prior cycles. All three can reference earlier propositions through their :about roles, and the architectural symmetry between them is exactly this: each points at the cognitive flow the agent has already produced to ground what it renders.
Consider an agent that has lifted observations about a nearby vehicle, an evaluation of risk, and a decision to wait. These are existing propositions in the subject stream, produced by the cognitive flow of Observe, Orient, and Decide:
(:distance (:as .f2 :in .S1)
:object :Car#1212
:meters 12.5)
(:risk (:as .e1 :in .S1)
:level :high
:event .f2)
(:decision (:as .d1 :in .S1)
:agent :Self
:action :wait)
A reflective proposition produced at some step boundary may surface as reflection:
(:reflection (:as .r1 :in .S1)
:agent :Self
:about [:set .f2 .e1 .d1]
:content (:assessment
:subject :Car#1212
:judgment :unsafe_to_cross)
:utterance "The car is too close to cross safely.")
A communicative action realized in the cycle’s Act phase may externalize a corresponding interpretation to a passenger:
(:speak (:as .sp1 :in .S1)
:agent :Self
:to :Passenger
:type :warning
:about [:set .f2 .e1 .d1]
:content (:assessment
:subject :Car#1212
:judgment :unsafe_to_cross)
:utterance "Hold on - that car is too close.")
Both propositions reference the same underlying cognitive flow (.f2, .e1, .d1) through their :about roles, and both carry an :assessment as their :content - a propositional payload structured in the same notation as the rest of the agent’s cognitive flow. Each also carries an :utterance rendering that payload as a natural-language string, one addressed to :Self and one delivered to the passenger. The cognitive flow is .f2, .e1, .d1 themselves - the perception, the evaluation, the decision - and it exists in the stream whether or not any verbal rendering accompanies it.
The cognitive grounding of speech acts is direct: a :speak references propositions from the stream itself through its :about chain, reaching back to the perception, evaluation, and decision propositions that prior cognitive flow produced. This is what makes the agent’s communicative behavior grounded rather than confabulated - the speech act is traceable to the actual stream content that motivated it, not to a confabulated rationalization. Reflection and internal speech, when present, are additional stream content the :speak may also reference, but they are not required for grounding; the traceability runs through the :about chain directly into the cognitive flow.
4.3 Reflection’s Within-Cycle and Across-Cycle Influence
A consequence of OODA’s phase sequencing is that the temporal effect of reflection depends on which phase of the cycle produced it. OODA phases consume what the upstream phase delivered: Orient processes what Observe made available, Decide processes what Orient produced, Act executes what Decide committed to. Downstream phases do not re-read the stream mid-cycle. Reflection follows this same sequencing as a natural consequence of being stream content.
Reflection produced in the Observe phase. Steps in Observe complete before Orient begins. Any reflective proposition produced at an Observe-phase step boundary is in the stream when Orient runs, and shapes the situation Orient constructs and therefore the commitment Decide selects. Observe-phase reflection has within-cycle influence.
Reflection produced in the Decide or Act phases. Steps in these phases complete after Orient has already constructed its situation. Reflection produced at these boundaries enters the stream, but the phase that would have used it has already run. It is present in the stream when the cycle ends and enters the next cycle’s Observe phase through internal sensing (Paper 06, Section 8). Decide- and Act-phase reflection has across-cycle influence.
The Orient phase sits between these two cases. Orient is itself a multi-step process, and a reflective lift produced at an early Orient step boundary may still be available to later Orient steps in the same cycle. Reflection produced at the end of Orient, however, arrives too late to shape the orientation that would have used it, and influences only the next cycle. The general principle holds: reflection has within-cycle influence when it lands in the stream before the step that consumes it, and across-cycle influence otherwise.
This distinction sharpens the analysis of cycles that combine reflection with internal or external speech. A cycle in which Observe-phase :reflection precedes an Act-phase :speak - whether self-directed or other-directed - is one where the reflection was visible to the situation Decide selected over, and is causally upstream of the speech act. A cycle in which :reflection was produced only after Orient is one where the reflection occurred but did not shape this cycle’s action - it will instead shape what the next cycle perceives.
The within-cycle and across-cycle distinction describes where reflection lands in time, not two separate modes of action realization. OAT’s architecture supports smooth realization of an action across any number of cycles - one cycle when the journey from recognition to commitment to execution compresses into a single loop, many cycles when the realization unfolds across an extended sequence of loops (§4.1.1). Reflection that arises in any of those cycles enters the stream and contributes to whichever cycle’s downstream phases can read it. Whether that cycle is the present one or a subsequent one is a matter of where the reflection landed, not a difference in kind. Single-cycle and multi-cycle action realization are both first-class behaviors of the same loop, and the cycle architecture is what makes both smooth.
4.4 Traceability and Transparency
Every :speak is connected to the cognitive flow that produced it through its :about references. The agent can explain why it said something by following those references back to the cognitive flow that grounded the utterance. The reference chains established in Paper 06 (Section 2.2.4) make this traceability architectural rather than reconstructive: the connections exist in the stream itself, not generated post-hoc when an explanation is requested.
When reflection is present, it adds a second layer of visibility: the agent’s reflective propositions show how it interpreted its lifted contents to itself, in addition to how it externalized them to others. Both kinds of propositions sit in the same stream, both reference the same cognitive flow, and both remain available to subsequent cycles and critic mechanisms.
5. Explicit Dialogue Structure at the Propositional Layer
5.1 Beyond Surface Patterns
Most systems that handle conversation operate on surface patterns. Template-based systems match input patterns to response templates. Large language models produce contextually appropriate responses through learned statistical regularities. In both cases, dialogue structure - the causal and functional relationships between utterances - remains implicit. Practitioners of conversational design have long observed that workable dialogue systems require attention to the structural properties of conversation itself - turn-taking, repair, sequencing, common ground [11] - rather than treating each utterance as an isolated input-output pair.
An OAT agent represents dialogue structure explicitly at the propositional layer. When the agent receives an utterance, it does not merely process the content. It classifies the speech act, represents its relationship to prior speech acts, and models how it affects the current cognitive state.
5.2 Propositional Representation of Speech Acts
Using the propositional representation language introduced in Paper 06, speech acts and their relationships can be represented as explicit, addressable structures.
Because Paper 10 is the first place where worked examples span multiple OODA cycles, it is also the first place where the propositional language’s cross-cycle scoping convention needs to be made explicit. Variables introduced in one cycle (.sa1, .o1, etc.) are local to that cycle for readability. When a later cycle references a proposition lifted in an earlier cycle, the reference is qualified with the originating situation tag: .S1.sa1 from cycle 2 refers to the proposition lifted as .sa1 in cycle 1. These local variables and the scoping convention are notational sugar for human authoring and reading; the underlying stream uses identifiers that are unique within the agent’s stream (something like :SpeechAct#849584958) that the local notation resolves to.
Consider a scenario where an agent is engaged in problem-solving and receives an interruption. The agent’s response unfolds across two cycles: first a decision cycle in which the agent switches tasks, then a communication cycle in which the agent reports the switch to the user.
;; CYCLE 1 (.S1) - decision cycle, no external speech
;; OBSERVE - agent's current planning state (sensed internally)
(:planning (:as .p1 :in .S1)
:agent :Self
:task :DataMigration
:status :in_progress)
;; OBSERVE - incoming speech act, linguistic surface
(:speech_act (:as .sa1 :in .S1)
:speaker :User
:utterance "The production database is down, can you look at it right away?")
;; ORIENT - propositional interpretation of the incoming speech act
(:interpretation (:as .i1 :in .S1)
:of .sa1
:type :request
:urgency :high
:content (:fix_bug
:system :ProductionDB
:severity :critical))
;; ORIENT - relationship to current plan
(:interrupts (:as .r1 :in .S1)
:event .i1
:disrupts .p1)
;; ORIENT - urgency assessment
(:assessment (:as .e1 :in .S1)
:event .i1
:urgency_confirmed :yes
:reason (:production_system_at_risk))
;; DECIDE - cognitive commitment to a future action
(:decide (:as .d1 :in .S1)
:agent :Self
:content (:switch_task
:from .p1
:to (:fix_bug :system :ProductionDB)))
;; ACT - the action executes; lifted as an action record
(:switch_task (:as .a1 :in .S1)
:agent :Self
:from .p1
:to (:fix_bug :system :ProductionDB)
:about [:set .i1 .r1 .e1])
The decision .d1 records the agent’s commitment to switching tasks; the action record .a1 lifts the executed switch into the stream. No speech act has occurred yet - the user is still waiting. Both .d1 and .a1 persist in the stream and become available to subsequent cycles.
;; CYCLE 2 (.S2) - communication cycle
;; All .S1 propositions remain accessible via internal sensing.
;; OBSERVE - new observations in .S2
(:unanswered (:as .o1 :in .S2)
:speech_act .S1.sa1)
(:recent_action (:as .o2 :in .S2)
:action .S1.a1)
;; DECIDE - cognitive commitment to a speech act
(:decide (:as .d2 :in .S2)
:agent :Self
:content (:speak
:to :User
:type :report
:content (:report
:event (:task_switched
:to (:fix_bug :system :ProductionDB)))
:utterance "I've switched to the production issue."))
;; ACT - the speech act executes; lifted as an action record
(:speak (:as .a2 :in .S2)
:agent :Self
:to :User
:type :report
:about [:set .o1 .o2]
:in_response_to .S1.sa1
:content (:report
:event (:task_switched
:to (:fix_bug :system :ProductionDB)))
:utterance "I've switched to the production issue.")
Cycle 2’s action record .a2 is a speech act reporting the completed switch. Its :content is a :report proposition describing what happened in cycle 1 - not a promise about future action, but a statement about what has already occurred.
Each cycle produces three kinds of stream content visible in the example: the orientation work that builds the situation, the :decide proposition that records a cognitive commitment to a future action, and the action record that lifts the executed action into the stream during Act. The decision and the action record mirror each other structurally - the decision is what the agent committed to, the action record is what actually happened - and future cycles can reference either, depending on whether they care about the cognitive commitment or the causal execution.
The action record .a2 carries two distinct kinds of references that should not be conflated. The :about role points at the agent’s own cognitive grounding for the utterance - in this example, the .S2-tagged observations .o1 and .o2 that together explain why the agent is saying this. The :in_response_to role points at the incoming utterance that occasioned the response - .sa1, the user’s original question, which was lifted in .S1 and remains stably referenceable. The first relationship is cognitive: it traces the speech act back to the current cycle’s reasoning. The second relationship is conversational: it ties the speech act into the dialogue structure as a reply, regardless of when the prior turn was lifted. Both are meaningful and both are represented explicitly, but they answer different questions and are kept separate at the propositional layer.
The example also illustrates how prior propositions are referenced across cycles. In cycle 2, the prior subject stream from .S1 is exposed through internal sensing (Paper 06, Section 8). When cycle 2 needs to reference a proposition that was lifted in cycle 1, the reference is qualified: .S1.sa1 and .S1.a1 rather than the bare .sa1 and .a1 used while still inside cycle 1. The qualification makes the originating cycle explicit and prevents collision with cycle 2’s own local variables (which can freely reuse short names like .o1, .o2). The cycle 2 action record grounds itself in the new .S2 observations through :about [:set .o1 .o2], while :in_response_to .S1.sa1 crosses cycles directly because dialogue structure is about speech-act-to-speech-act relationships regardless of which cycle observed what.
5.3 Dialogue Patterns as Cognitive Patterns
Because speech acts are embedded in the OODA loop, dialogue patterns map to cognitive patterns:
Follow-up occurs when a prior speech act generates expectations that are tracked across cycles. The agent perceives an unanswered question as an open commitment and generates a follow-up speech act when the commitment is not resolved.
Correction occurs when new observation contradicts a prior assertion. The agent detects the inconsistency during Orient, and the correction speech act emerges from the agent’s own error-detection process - not from a dialogue rule.
Escalation occurs when repeated speech acts fail to produce the expected effect. The agent models the accumulating failure across cycles, and the escalation pattern (from “Can you help?” to “I need help now” to “This is critical”) reflects increasing urgency in the agent’s evaluation, not a scripted progression.
Negotiation occurs when agents exchange speech acts that iteratively constrain the space of acceptable outcomes. Each proposal is an action candidate evaluated through ATTM-weighted future assessment (Paper 07), and counter-proposals emerge from the same evaluative machinery.
These patterns are not stored as templates. They emerge from the agent’s cognitive process operating over explicit propositional representations of communicative events and their relationships.
6. Communication and Attachment-Based Evaluation
6.1 Speech Acts Carry ATTM Weight
Paper 07 introduced the Attachment Matrix (ATTM) as the mechanism by which agents evaluate actions in terms of their consequences for attached entities across time horizons. Speech acts, as genuine actions, are subject to the same evaluation.
A promise made to an entity with high ATTM weight carries greater evaluative significance than the same promise to an entity with low weight. A refusal directed at a strongly attached entity generates higher penalty signals. A request for help, when directed to the right entity, reflects the agent’s model of which attachments can absorb the cost of assistance.
This means that the agent’s communicative behavior is not uniform across the entities it interacts with. It is shaped by the same attachment structure that shapes all other action selection. The agent speaks differently to different entities - not through stylistic variation, but because the evaluation of communicative acts differs based on ATTM weights.
ATTM enters the architecture at Decide, where candidate actions are evaluated against attachment-weighted projected trajectories. Both internal and external speech are committed actions, so both are subject to this evaluation - a :speak :to :Self no less than a :speak :to :User (or, in Paper 04’s more detailed terms, candidate bundles in which a speech act may be primary or subordinate). Reflection is not an action and does not pass through Decide’s evaluation. What ATTM shapes indirectly is which action candidates cognitive flow proposes during Orient and which ones Decide commits to.
6.2 Communicative Acts as Well-Being Interventions
Speech acts modify predicted well-being trajectories. Reassurance stabilizes predicted near-term well-being of an attached entity. Warnings protect long-horizon well-being by enabling preventive action. Difficult assessments delivered carefully can preserve long-term trust at the cost of short-term comfort.
The agent evaluates candidate speech acts against ensembles of possible futures (Paper 07, Section 5.3) - what happens if I say this vs. that vs. nothing? Silence is always a candidate, evaluated for its consequences just as any other action.
Behaviors associated with ethical maturity - hesitation before delivering bad news, careful framing of difficult assessments, willingness to sacrifice short-term comfort for long-term trust - emerge from this attachment-weighted evaluation rather than from applied rules.
6.3 Critics Over Communicative Structure
The propositional representation of dialogue structure developed in §5 - :speak, :about, :content, :utterance, :in_response_to - makes specific communicative failure modes architecturally visible. An ungrounded :speak has no :about references pointing back at prior cognitive flow. An unresponsive reply carries an :in_response_to pointing at a speech act whose :about chain does not intersect with its own. These conditions are detectable as structural mismatches at the propositional layer, independent of natural-language surface analysis, and can be picked up by critics in the sense of Paper 09. A fuller treatment of communication-specific critics is left to future work.
7. Prior Implementation and Validation
7.1 Pre-LLM Speech Act Architecture
Before 2024, the author experimented with agent systems in which speech acts were deeply integrated with the agent’s planning, problem-solving, and action-selection processes. These systems used NLP templates, intent recognizers, and slot fillers - the available implementation technology of the time - to realize an architecture in which communication was not handled by a separate dialogue manager but emerged from the agent’s cognitive process.
The speech act taxonomy in these systems was extensive and directly tied to cognitive state. Acts such as PROBLEM SOLVING STEP NOTIFICATION (with graded variants), CANNOT SOLVE PROBLEM NOTIFICATION (also graded), SOLVED PROBLEM NOTIFICATION, and PROBLEM SOLVING STEP FAILED, NOT ENOUGH KNOWLEDGE were not dialogue moves - they were communicative surfaces of specific cognitive states. The agent was not following a conversational script. It was reporting its own problem-solving process. In the pre-OAT terminology of those systems, the cognitive state was the primary thing and the speech act was its outward surface; OAT’s propositional layer, developed later, provides the explicit reference machinery (§4’s :about chains) that those systems were approximating through direct template-to-state bindings.
Dialogue sequencing was similarly driven by cognitive structure rather than conversational scripts. When the agent’s problem-solving process branched into a sub-task, the communicative output followed: the agent would announce the sub-task, report progress within it, and then resume reporting on the parent task. When the agent switched to a different goal - because of an interruption, a dependency, or a re-prioritization - the speech act sequence shifted accordingly. The conversation was not managed by a dialogue planner operating independently of the agent’s reasoning. It was generated by the reasoning process itself, with each speech act reflecting a specific transition in the agent’s planning or execution state.
7.2 What Changed and What Didn’t
The advent of large language models made the NLP implementation layer - templates, intent recognizers, slot fillers - largely obsolete. An LLM can recognize speech act types, generate contextually appropriate responses, and handle natural language variation far more flexibly than engineered NLP pipelines.
What did not change is the architectural principle. The integration of communication with cognitive process - the claim that speech acts are grounded in and traceable to the agent’s ongoing OODA activity - is independent of the implementation technology. The templates were an approximation of what the propositional lift now provides formally: explicit, referenceable stream contents that speech acts can point at through :about chains, making the grounding architectural rather than implicit in a template-to-state binding.
The earlier work was a valuable experiment in re-thinking communication as part of more general agent behavior rather than as a separate subsystem. Communication was integrated directly with problem-solving, planning, and action selection - the agent’s speech acts were generated by the same processes that drove its non-communicative behavior. The NLP was basic, but the integration was sound. What OAT adds is the architectural machinery that makes the integration formally precise: the propositional layer with explicit :about references (§4, Paper 06), the three-stage pipeline from cognitive flow to committed speech act to rendered utterance (§4.1.1), and critics that operate over the resulting propositional structure (§6.3).
8. Communication as OODA Coupling
Communication connects agents. When agents communicate, their OODA loops become coupled: one agent’s actions become another’s observations, interpretations align through interaction, decisions are coordinated. This coupling operates across fundamentally different time scales, and the distinction between these scales has consequences for how agency persists.
8.1 Real-Time Coupling
When two OAT agents communicate in real time, their OODA loops enter a synchronous coupling. One agent’s speech act, executed during Act, enters the other agent’s Observe phase. The receiving agent classifies it, orients to it, and generates a response through its own OODA cycle. The response re-enters the first agent’s observation.
This coupling produces shared attention, alignment of interpretation, and coordinated action. It is not merely information exchange. Because speech acts create commitments (§3.2), each communicative exchange modifies the commitment landscape of both agents. A negotiation between two OAT agents is a sequence of mutually constraining speech acts, each evaluated through the respective agents’ ATTM structures and each creating commitments that persist across cycles.
The stability and quality of real-time coordination depend on whether commitments and dialogue structure are modeled explicitly - agents that represent the causal relationships between speech acts can detect and repair coordination failures that agents operating on latent patterns cannot.
8.2 Persistent Coupling Through Stories
Speech couples closures in the present. Stories couple them across time.
Agents can produce and consume stories - structured narratives that preserve interpreted experience in a form that other agents, or the same agent at a later time, can ingest during Orient. Stories take many forms. Experiential records are an agent’s account of a situation it encountered, how it oriented, what it decided, and what happened - a compressed cognitive episode preserved as lifted propositions. Generalizations are stories constructed by abstracting across multiple episodes, capturing patterns rather than individual events - “when this kind of situation arises, these are the relevant considerations.” Normative narratives encode evaluative structure explicitly - what matters, what boundaries must be respected, what trade-offs are acceptable. A cautionary tale transmits a guardrail. A case study transmits an ATTM weighting - what was worth sacrificing for, and why. Many other forms exist - myths, instructions, autobiographies, scientific reports, parables, scripts - each preserving cognitive structure in a way other agents can re-ingest.
OAT treats the construction and interpretation of stories as a central cognitive capability, not an auxiliary one. The situational propositional language developed across the series - situations as first-class addressable entities, lifted propositions, qualified cross-situation references - is well-suited to representing stories of essentially any kind. An agent can use the same propositional machinery to represent its own past episodes, another agent’s reported experience, a historical account, a friend’s anecdote, a TV show, a film, a novel, or a fictional scenario it is constructing for its own purposes. In each case, the agent constructs a propositional representation of a situated activity - characters as referenced entities, their apparent actions as lifted propositions, their inferred reasoning as :about chains - and this representation enters the stream where it becomes available to subsequent cognition. Stories are not passive records. They are mechanisms through which agentic structures persist beyond any single interaction or any single agent’s operational lifetime. Winston’s thesis that storytelling is central to human intelligence rather than peripheral to it [12] supports this architectural claim: the capacity to construct and comprehend stories is not a communication skill but a cognitive capability that enables experience to be preserved, transmitted, and applied beyond its original context.
This connects directly to the cultural transmission of ATTM structures described in Paper 07. An accumulated store of stories - experiential records, generalizations, normative narratives, institutional practices - carries attachment weights, guardrails, and evaluative patterns across agents and across time. The propositional lift provides the mechanism by which these transmitted structures can be explicitly represented and integrated into an agent’s evaluative architecture rather than remaining latent influences.
8.3 Coupling and the Extension of Agency
Real-time coupling and persistent coupling through stories together enable forms of agency that reach beyond what any single closure can sustain alone. Individual closures are limited by bounded memory, finite lifespan, and partial perception. Through communicative coupling, these limits can be partially overcome: perception is shared across closures, orientation is informed by stories from agents the current one never met, commitments are coordinated in real time, and accumulated experience persists across generations through transmitted stories and ATTM-bearing narratives (§8.2, Paper 07).
What makes this possible is OAT’s deep integration of communication with the agent’s cognitive architecture: speech acts are generated from the same control process that drives all behavior, grounded in the same cognitive flow, evaluated through the same attachment-weighted mechanism, and monitored by the same critic architecture. This integration is what enables coupling to reach into the evaluative and commitment-forming processes of each closure, producing coordination that is robust, traceable, and persistent.
Each closure participating in a coupled interaction remains its own closure - its own OODA loop, subject stream, ATTM, and cycle rate. Coupling is a relationship between closures, not a merger. Whether the emergent coordination patterns that arise between closures themselves constitute closures at a higher scale is left to future work.
9. Conclusion
Communication is not a capability bolted onto agency. It is an extension of the OODA loop itself - new observation channels, new action modalities, and new commitment structures, causally integrated with orientation, evaluation, and decision across every phase of the control architecture.
Within OODA Agency Theory, speech acts are actions in the full sense developed by Paper 04 and Paper 02: a :speak may serve as the primary commitment of the cycle’s action bundle or as a subordinate action accompanying a non-communicative primary commitment. Its cognitive grounding - traceable through :about references in the propositional layer - draws on contents lifted from across the whole cycle and from prior cycles.
Cognitive flow is the primary cognitive activity of the loop, and reflection, internal speech, and external speech are three forms through which some of that cognitive flow can surface in linguistic form. Reflection is substrate verbalization that enters the stream without being committed as an action. Internal and external speech are committed actions, distinguished only by recipient: a :speak :to :Self and a :speak :to <other> go through the same candidate-construction, commitment, and execution pipeline, both carrying :content as a propositional payload and :utterance as its natural-language rendering. All three reference the cognitive flow they ground in through their :about roles, which is what makes the agent’s communicative behavior cognitively grounded rather than confabulated.
By modeling dialogue structure explicitly at the propositional layer, OAT agents maintain a full causal picture of communicative interaction. The :about role traces a speech act back to its cognitive grounding; the :in_response_to role ties it to the conversational structure as a reply. These two relationships are kept distinct, and together they make the communicative behavior of an OAT agent traceable along both the cognitive and the dialogue dimensions.
The architectural principle - communication embedded in cognition, not layered on top - was independently validated in pre-LLM implementations that integrated speech acts directly with problem-solving, planning, and action selection. What OAT adds is the architectural machinery those earlier systems lacked: the propositional layer with explicit :about references and dialogue roles, making communicative grounding formal rather than implicit in template-to-state bindings.
Speech couples closures in the present; stories couple them across time. Through this coupling, agency persists beyond what any single closure can sustain alone.
References
[1] Austin, J. L. (1962). How to Do Things with Words. Oxford University Press.
[2] Searle, J. R. (1969). Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.
[3] Tomasello, M. (1999). The Cultural Origins of Human Cognition. Harvard University Press.
[4] Tomasello, M. (2003). Constructing a Language: A Usage-Based Theory of Language Acquisition. Harvard University Press.
[5] Kamp, H., & Reyle, U. (1993). From Discourse to Logic. Kluwer Academic Publishers.
[6] Nirenburg, S., & Raskin, V. (2004). Ontological Semantics. MIT Press.
[7] Jurafsky, D., & Martin, J. H. (2009). Speech and Language Processing. 2nd edition. Prentice Hall.
[8] Boyd, J. R. (1996). The Essence of Winning and Losing. Unpublished briefing.
[9] Newell, A. (1982). The Knowledge Level. Artificial Intelligence, 18(1), 87-127.
[10] Vygotsky, L. S. (1986). Thought and Language. MIT Press.
[11] Moore, R. J., & Arar, R. (2019). Conversational UX Design: A Practitioner’s Guide to the Natural Conversation Framework. ACM Books.
[12] Winston, P. H. (2011). The Strong Story Hypothesis and the Directed Perception Hypothesis. AAAI Fall Symposium Series.
[13] Bordini, R. H., Hübner, J. F., & Wooldridge, M. (2007). Programming Multi-Agent Systems in AgentSpeak using Jason. John Wiley & Sons.
[14] Bellifemine, F. L., Caire, G., & Greenwood, D. (2007). Developing Multi-Agent Systems with JADE. John Wiley & Sons.
[15] Rao, A. S., & Georgeff, M. P. (1995). BDI Agents: From Theory to Practice. Proceedings of the First International Conference on Multi-Agent Systems (ICMAS-95).