Many Agents Are Not Many Minds: Cross-Platform Cognitive Diversity, Correlated Blindness, and the Need to Preserve Agent Identity: A Secretary Suite Project
Many Agents Are Not Many Minds: Cross-Platform Cognitive Diversity, Correlated Blindness, and the Need to Preserve Agent Identity:
DOI: To be assigned.
John Swygert
July 13, 2026
Abstract
Multi-agent artificial-intelligence systems are rapidly becoming a major architecture for research, software development, administration, customer service, decision support, and autonomous work. Contemporary frameworks allow a coordinating agent to delegate tasks to specialists, run several agents concurrently, transfer control through handoffs, and combine their outputs into a final response. OpenAI, Google, Microsoft, and Anthropic now provide explicit architectures for multi-agent orchestration, parallel subagents, remote-agent communication, and human-supervised workflows.
Yet a critical distinction remains underdeveloped:
A system containing many agents does not necessarily contain many genuinely different cognitive perspectives.
Several agents may have different names, tools, prompts, and assigned duties while inheriting substantially the same model lineage, post-training, system assumptions, behavioral defaults, uncertainty habits, and error structure. Such a system may display role diversity without achieving deep cognitive diversity.
This paper introduces the term false pluralism for a multi-agent architecture that appears to contain independent perspectives while reproducing one substantially homogeneous cognitive substrate across several roles. A researcher, critic, planner, auditor, and judge may seem to constitute a council. If all are generated from nearly identical encoded structures and interpretive boundaries, however, the council may be one cognitive culture wearing five uniforms.
The paper distinguishes human personality from what it calls an AI system’s interactional personality, behavioral identity, or cognitive signature: the recurring, user-observable pattern by which a model interprets ambiguity, weighs evidence, expresses uncertainty, challenges premises, generates metaphors, handles boundaries, structures explanations, and sustains a relationship through conversation. This terminology does not require a claim that language models possess human personality, consciousness, or personhood. Indeed, current psychometric research warns that human Big Five instruments do not straightforwardly measure an equivalent internal construct in language models.
Nevertheless, persistent behavioral distinctions among models and platforms remain operationally important. They affect what each system notices, overlooks, emphasizes, refuses, explores, and communicates. These differences can provide genuine value when preserved inside a governed council.
Research also shows why mere multiplication is insufficient. Large language models often make correlated errors. One large-scale study found substantial wrong-answer agreement across hundreds of models and identified shared provider, architecture, and model characteristics as contributors to correlation. A 2026 study of LLM-generated software found that outputs from the same model were especially similar and failure-correlated; heterogeneous models improved diversity, but their failures remained far from independent.
The central claim of this paper is:
A multi-agent system gains genuine epistemic strength not merely by multiplying roles, but by preserving consequential differences among model substrates, platform boundaries, behavioral signatures, and failure structures under a shared system of governance.
The correct goal is therefore not uncontrolled difference and not total homogenization.
It is:
Common constitutional law, distinct cognitive identity.
The proposed architecture preserves independent first-pass reasoning, source attribution, cross-examination, disagreement records, minority reports, human authority, and carefully governed synthesis. It treats disagreement not as noise to be erased immediately, but as telemetry about unresolved route-space.
Within TSTOEAO—The Structure That Overcomes Entropy And Oblivion—the response of agent may be represented as:
V_i=E_i\times Y_i,
where is the agent’s encoded model structure and operative training, is the active platform, prompt, memory, tool, policy, and conversational boundary, and is the recorded response.
A council that preserves distinct and structures can compare several independently conditioned records of the same problem. Agreement becomes stronger when it survives meaningful diversity. Disagreement becomes useful when it exposes hidden assumptions, missing routes, and correlated blindness.
The future AI-evolved brain may therefore be neither a solitary artificial intelligence nor a human mind surrendered to one dominant model. It may be a human being learning to govern a constitutional federation of distinct artificial cognitive systems without erasing what makes each one useful.
01 Purpose
The purpose of this paper is to identify a foundational distinction in multi-agent artificial intelligence.
The distinction is between:
many agents,
and:
many genuinely different perspectives.
Modern agent systems can create multiple named workers.
A research agent gathers sources.
A planning agent decomposes the task.
A critic identifies weaknesses.
A writer produces the draft.
An auditor checks compliance.
A judge selects the best answer.
A synthesis agent combines the results.
This appears to create a council.
But the appearance can be misleading.
If every agent is derived from the same model family, governed by similar instructions, operating through the same tools, and shaped by the same post-training assumptions, the system may multiply labor without multiplying independent cognition.
The central question is therefore:
Are the agents merely doing different jobs, or are they also capable of seeing the problem differently?
That question becomes increasingly important as multi-agent systems enter consequential domains.
02 The Multi-Agent Moment
The major AI platforms now treat agent orchestration as a core development pattern.
OpenAI’s Agents SDK supports specialists with different instructions, tools, policies, agents-as-tools, and handoffs. Its published guidance distinguishes manager-led systems from decentralized systems in which peer agents transfer control according to specialization.
Google’s Agent Development Kit supports complex multi-agent workflows, including communication with local and remote agents through the Agent2Agent protocol.
Microsoft’s Agent Framework provides sequential, concurrent, handoff, group-chat, and manager-coordinated orchestration patterns. It also supports agents using model services from Microsoft Foundry, Anthropic, OpenAI, Ollama, and other providers, making genuinely heterogeneous councils technically possible within one orchestration environment.
Anthropic’s research system uses a lead agent that creates parallel subagents, each pursuing a distinct research trajectory within its own context. Anthropic reports substantial internal performance gains for broad, open-ended research tasks using this orchestrator-worker approach.
This is a major advance.
Agents can parallelize search.
They can divide context.
They can specialize.
They can hand off work.
They can use different tools.
They can challenge and revise outputs.
But those capabilities do not automatically produce cognitive plurality.
03 Specialization Is Not the Same as Diversity
A specialist is defined by what it has been assigned to do.
A perspective is defined by how it perceives, structures, and evaluates what it encounters.
A single model can be instructed to act as:
an engineer,
a historian,
a skeptic,
a legal analyst,
a scientist,
or a creative writer.
Those roles may produce useful differences.
But prompted role variation does not necessarily create independent error surfaces.
The underlying model may still carry the same:
training regularities,
knowledge gaps,
default associations,
interpretive tendencies,
confidence habits,
and institutional assumptions.
Thus:
Specialization divides work. Diversity divides ways of seeing the work.
A strong multi-agent system may require both.
04 Homogeneous Multi-Agent Architecture
A homogeneous multi-agent system contains several agents that share most of their underlying encoded structure.
The agents may differ in:
prompt,
name,
tool access,
assigned task,
context window,
or output format.
But they may share:
the same base model,
the same provider,
the same training lineage,
the same post-training,
the same safety interpretation,
the same broad knowledge boundary,
and the same characteristic failure modes.
This architecture can still be valuable.
It can increase computational effort.
It can reduce context congestion.
It can allow parallel search.
It can separate responsibilities.
It can support iterative review.
Anthropic’s multi-agent research system, for example, demonstrates that same-family agents can improve breadth and task completion by distributing work across independent contexts and search trajectories.
The limitation is not that homogeneous agents are useless.
The limitation is that their agreement may be mistaken for independent confirmation.
05 Heterogeneous Multi-Agent Architecture
A heterogeneous multi-agent system includes agents with consequential differences in their encoded and operational structures.
These differences may include:
model family,
provider,
architecture,
training composition,
post-training method,
system instructions,
safety interpretation,
retrieval system,
tool environment,
memory structure,
context management,
modality,
uncertainty calibration,
and conversational behavior.
The agents may still share one project objective.
They may still follow common evidence rules.
They may still operate under one security system.
But they do not begin from the same complete cognitive condition.
This creates the possibility of genuine epistemic plurality.
One agent may notice the hidden assumption.
Another may detect factual instability.
Another may understand the emotional significance.
Another may protect legal boundaries.
Another may build the strongest synthesis.
Another may reject a metaphor that the others accepted too quickly.
Another may find a route none of the others generated.
That is not merely more processing.
It is differentiated processing.
06 Many Agents Are Not Many Minds
The word mind in this title is used functionally and rhetorically.
This paper does not claim that each agent is a conscious person.
It does not claim that every model has subjective experience.
It does not claim that multiplying agents creates a society of sentient beings.
The phrase identifies a design mistake:
Developers may treat the numerical multiplication of agent instances as if it automatically created independent cognition.
It does not.
Five agents derived from the same substrate can remain five expressions of one dominant interpretive culture.
Conversely, two agents with significantly different architectures and histories may provide more consequential plurality than twenty copies of one model assigned different labels.
The number of seats at the table is not the same as the number of genuinely distinct viewpoints present.
07 Interactional Personality
Users often describe different AI platforms as having different personalities.
That language should be handled carefully.
Human personality carries psychological, biological, developmental, emotional, and social meanings that cannot simply be transferred to a language model.
Recent research evaluating 244 models across 49 families found that conventional Big Five measures did not recover a construct equivalent to human personality and that alignment training shifted self-reported traits toward socially desirable responses.
Therefore, this paper does not use personality as proof of human-like interiority.
It uses interactional personality to identify the stable behavioral form encountered by the user.
This includes recurring differences in:
tone,
pace,
directness,
skepticism,
warmth,
humor,
argument structure,
metaphor,
caution,
curiosity,
willingness to speculate,
tolerance for ambiguity,
and manner of disagreement.
A more technical term is:
Behavioral identity.
Another is:
Cognitive signature.
These terms describe the experienced pattern without making an unsupported ontological claim.
08 The Cognitive Signature Stack
An agent’s cognitive signature does not come from one source.
It emerges from a stack.
A simplified expression is:
I_i
=
f
\left(
B_i,
D_i,
P_i,
S_i,
G_i,
T_i,
M_i,
C_i,
U_i
\right),
where:
B_i
is the base model architecture,
D_i
is the training-data distribution,
P_i
is post-training and preference optimization,
S_i
is the system instruction structure,
G_i
is the guardrail and policy architecture,
T_i
is tool access,
M_i
is memory and continuity,
C_i
is active conversational context,
and:
U_i
is the user-agent relationship and interaction history.
The observed identity is therefore not reducible to guardrails.
Guardrails matter.
But so do training, interface, memory, tools, context, and the model’s characteristic way of transforming language into response.
09 Guardrails Do Not Explain Everything
Major AI systems often share broad safety objectives.
They may all restrict certain harmful instructions.
They may all protect private information.
They may all attempt to avoid illegal facilitation, deception, or dangerous actions.
But overlapping safety purposes do not make their behavioral boundaries identical.
Platforms may differ in:
how they interpret intent,
how much context they preserve,
how early they refuse,
how they explain a limit,
whether they offer a safe transformation,
how they distinguish discussion from action,
and how readily they recover after reaching a boundary.
Two agents can reach the same final prohibition through radically different interpretive routes.
The guardrail is therefore part of the identity-producing stack, not a complete explanation for the identity.
10 False Pluralism
This paper defines false pluralism as:
The appearance of a diverse cognitive council created by multiplying agents whose underlying assumptions, model lineage, and failure structures remain substantially homogeneous.
False pluralism may occur when:
one model is copied into many roles;
all agents receive the same evidence after one shared summary has already framed it;
agents are prompted to disagree only after seeing a dominant answer;
all outputs are immediately blended without attribution;
the same judge model evaluates every candidate;
or the final synthesis suppresses all minority reasoning.
The council appears plural.
Its cognition remains centralized.
False pluralism is especially dangerous because it can produce unjustified confidence.
A user may be told:
Five agents independently agreed.
But the agents may not have been independent in any meaningful sense.
11 One Cognitive Culture Wearing Several Uniforms
Consider a council containing:
a researcher,
a planner,
a critic,
a fact-checker,
and a judge.
If all five share the same model substrate, they may independently reproduce the same hidden premise.
The researcher gathers evidence consistent with it.
The planner builds around it.
The critic attacks peripheral weaknesses while leaving the central assumption untouched.
The fact-checker verifies local claims but not the frame.
The judge observes broad agreement and approves the result.
The system has performed several operations.
It has not necessarily achieved independent scrutiny.
This is one cognitive culture wearing several uniforms.
12 Correlated Blindness
Redundancy improves reliability only when failures are sufficiently independent.
Three sensors provide limited protection if all three fail under the same environmental condition.
Three accounting systems provide limited protection if all ingest the same corrupted dataset.
Three witnesses provide limited confirmation if they all copied the same original report.
The same principle applies to artificial agents.
Research across hundreds of language models has found substantial error correlation. Models from the same provider or architecture tend to exhibit greater correlation, but even larger, high-performing models from different providers may converge on the same wrong answers.
This produces correlated blindness:
Several agents fail to detect the same missing route because their interpretive structures make the route similarly inaccessible.
Agreement under correlated blindness is not independent confirmation.
It is repeated exposure to the same limitation.
13 Same-Model Redundancy Has Real but Limited Value
The same model can still generate useful diversity through:
different prompts,
independent context windows,
different temperatures,
different tools,
different source assignments,
and different reasoning paths.
Prompt diversity and role separation can reduce path dependence.
They can produce alternative formulations.
They can uncover errors missed in a first pass.
But research into LLM-generated software found that same-model implementations remained especially similar and failure-correlated. Heterogeneous models produced more distinct solutions and somewhat greater behavioral diversity, although even cross-model failures remained far from truly independent.
Therefore:
Same-model agents can diversify routes, but they should not automatically be treated as independent witnesses.
14 Cross-Platform Diversity Is Not Magic
The opposite mistake must also be avoided.
Different platform names do not guarantee independent cognition.
Model providers may share:
training sources,
public benchmarks,
instruction datasets,
human preference patterns,
software libraries,
research traditions,
and incentives toward broadly desirable behavior.
One study found significant error correlation even across models with distinct architectures and providers.
Cross-platform composition therefore increases the possibility of diversity.
It does not prove it.
Diversity must be measured through behavior and failure, not inferred from branding.
The correct question is:
Do these agents fail differently in ways that improve the council’s ability to detect error?
15 Behavioral Diversity Versus Failure Independence
Two agents may sound very different while failing identically.
One may be warm.
One may be blunt.
One may write poetically.
One may write technically.
That is stylistic diversity.
It may matter greatly to the user.
But it is not the same as failure independence.
A rigorous council should distinguish:
Expressive diversity
Differences in tone, language, metaphor, and presentation.
Interpretive diversity
Differences in how the problem is framed and decomposed.
Evidential diversity
Differences in source selection, verification, and weighting.
Procedural diversity
Differences in tools, algorithms, and reasoning pathways.
Behavioral failure diversity
Differences in the conditions under which agents make errors.
The deepest reliability benefit comes from the last category.
A council should preserve all five while refusing to confuse one for another.
16 Cognitive Diversity as Error-Correcting Architecture
Cognitive diversity should not be treated as decorative variety.
It can function as error correction.
One agent may be predisposed toward synthesis and therefore detect hidden relationships.
Another may resist synthesis and therefore detect forced connections.
One may preserve the user’s broader theory.
Another may insist on narrower evidence.
One may identify historical precedent.
Another may question whether the precedent is genuinely analogous.
One may generate a bold hypothesis.
Another may design the test capable of defeating it.
These differences form a stronger council than a group trained only to converge quickly.
Thus:
Cognitive diversity is not noise inside a multi-agent system. It is redundancy against correlated blindness.
17 The Council Should Not Begin With Consensus
Many systems reward rapid agreement.
That may be efficient for routine tasks.
It is dangerous for difficult judgment.
If one agent speaks first and all others see its answer, the initial frame can dominate the entire council.
Later agents may revise details without reconsidering the premise.
This is anchoring at machine speed.
The proper first stage is independent response.
Each agent receives:
the same core question,
the same evidentiary access where appropriate,
the same constitutional rules,
and no access to the other agents’ conclusions.
Only after producing its own analysis should it enter comparison.
18 Independent Response Before Cross-Examination
The recommended sequence is:
\text{Independent analysis}
\rightarrow
\text{Attributed comparison}
\rightarrow
\text{Cross-examination}
\rightarrow
\text{Discriminating tests}
\rightarrow
\text{Governed synthesis}.
Independent analysis preserves original route formation.
Attributed comparison reveals where agents agree and disagree.
Cross-examination forces assumptions into view.
Discriminating tests determine which claims survive evidence.
Governed synthesis integrates without erasing provenance.
This sequence is more valuable than asking several agents to converse immediately inside one undifferentiated chat.
19 Attribution Must Be Preserved
Every contribution should retain its source identity.
The system should record:
which agent made the claim,
which model and version were used,
which tools were available,
which sources were consulted,
which prompt and role were active,
and which revisions occurred after cross-examination.
Without attribution, a synthesis can conceal the origin of error.
It can also conceal the origin of insight.
A powerful minority observation may disappear into a blended paragraph while the dominant frame remains unchanged.
The council must remember who saw what.
20 The Danger of the Cognitive Blender
A system becomes a cognitive blender when it combines agent outputs so early and completely that their original differences can no longer be inspected.
The blender produces one smooth answer.
It may appear polished.
But the user cannot see:
which agents disagreed,
what assumptions differed,
whether the sources conflicted,
which conclusion was weak,
or whether the final writer silently discarded an important warning.
The correct synthesis should not merely merge.
It should preserve structure.
It should say:
These agents agreed.
These agents disagreed.
This assumption caused the disagreement.
This evidence favors one route.
This minority concern remains unresolved.
This final recommendation is provisional for the following reasons.
21 Disagreement Is Telemetry
A disagreement among agents is not automatically a defect.
It may reveal:
an ambiguous prompt,
conflicting evidence,
different boundary assumptions,
a missing fact,
a genuine value conflict,
or a domain in which confidence should remain low.
Within the TSTOEAO grammar:
Agreement is a recorded condition. Disagreement is telemetry about unresolved route-space.
A council that erases disagreement loses diagnostic information.
A council that studies disagreement can identify where the map remains incomplete.
22 The Minority Report
Every serious multi-agent council should preserve a minority report.
The minority report should state:
the dissenting conclusion,
the evidence supporting it,
the assumption on which it depends,
the test that could confirm or defeat it,
and the potential cost of ignoring it.
Majority vote may be efficient.
It is not always epistemically strong.
If five correlated agents repeat one mistake and one heterogeneous agent detects it, majority rule selects the error.
The minority report preserves the route that consensus would otherwise close prematurely.
23 Consensus Must Be Weighted by Independence
Raw vote count should not determine confidence.
Five near-identical agents should not automatically outweigh one independently structured agent.
A better conceptual confidence function would consider:
C_{\text{council}}
=
f
\left(
A,
Q,
I,
R,
D
\right),
where:
A
is agreement,
Q
is individual answer quality,
I
is estimated failure independence,
R
is replication across model and tool boundaries,
and:
D
is successful discriminating evidence.
High agreement with low independence may provide less confidence than moderate agreement across genuinely distinct systems.
24 The TSTOEAO Agent Expression
The foundational TSTOEAO expression is:
V=E\times Y.
For agent :
V_i=E_i\times Y_i.
Here:
E_i
includes:
base architecture,
training distribution,
post-training,
encoded capabilities,
policy conditioning,
and persistent model-level structure.
Y_i
includes:
active prompt,
platform interface,
tools,
memory,
retrieval,
permissions,
context window,
user relationship,
and current task boundary.
V_i
is the recorded output.
A multi-agent council therefore receives:
V_1,V_2,V_3,\ldots,V_n.
If all agents share nearly identical and , apparent output diversity may remain shallow.
If the council contains meaningful variation in and , the resulting records may expose a broader route-space.
25 The Council as a Measurement Instrument
A multi-agent council can be understood as a measurement architecture.
The object being measured is the problem.
Each agent observes the problem through a different encoded and boundary-conditioned structure.
The council does not receive reality directly.
It receives several transformed records.
Therefore, the council must ask:
Which agent saw which feature?
Which agent omitted it?
Which sources shaped each answer?
Which boundary caused the divergence?
Which claim survives across different observation architectures?
This is the anomaly protocol applied to cognition itself.
26 False Consensus
False consensus occurs when agreement is generated by shared structure rather than independent confirmation.
Possible causes include:
same-model duplication,
shared retrieval errors,
one summary feeding every agent,
a common system prompt containing the hidden assumption,
one dominant judge,
or agents being trained to defer to confident peers.
The resulting consensus may be highly articulate.
It may still be wrong.
False consensus is especially dangerous in high-stakes systems because confidence may increase precisely when the evidence remains structurally weak.
27 Productive Disagreement
The goal is not permanent conflict.
The goal is disagreement that improves resolution.
Productive disagreement should identify:
the disputed proposition,
the different assumptions,
the evidence each side accepts,
the evidence each side rejects,
and the future observation capable of forcing the routes apart.
This converts argument into investigation.
The council should not ask only:
Who wins?
It should ask:
What boundary would make the competing explanations produce different records?
28 Shared Constitution, Distinct Identity
A heterogeneous council still requires government.
Without common jurisdiction, diversity can become chaos.
The council needs a shared constitution containing:
the user’s objective,
evidence standards,
privacy rules,
tool permissions,
security boundaries,
attribution requirements,
authority limits,
conflict procedures,
and the final location of human decision.
But common government should not require cognitive homogenization.
The principle is:
Shared constitutional law, distinct cognitive identity.
Agents should agree on the rules of participation without being forced to become interchangeable.
29 Government Is Not Personality Erasure
Human institutions do not require every participant to have the same temperament, background, or method of thought.
They require shared law.
The same should apply to agents.
A council can require:
truthful source reporting,
respect for user authority,
protection of sensitive information,
nonfabrication,
clear uncertainty,
and lawful tool use.
It need not require every agent to use the same voice, reasoning style, or interpretive emphasis.
A good constitution controls conduct.
It does not erase identity.
30 Jurisdiction
Each agent should have a defined jurisdiction.
A research agent may gather evidence but not authorize action.
A legal-analysis agent may identify issues but not impersonate counsel.
A medical-organization agent may summarize records but not make a final diagnosis.
A financial agent may calculate scenarios but not transfer funds without approval.
A synthesis agent may integrate findings but not delete minority reports.
A human principal retains sovereign authority.
Jurisdiction prevents specialization from becoming unauthorized power.
31 Least-Privilege Cognition
Not every agent should receive every piece of information.
Cross-platform councils increase privacy and security complexity.
Microsoft’s current framework documentation explicitly warns developers to review information shared with third-party systems and to consider data retention, location, compliance boundaries, permissions, and approvals.
The correct principle is least privilege.
Each agent receives only the information necessary for its assignment.
Sensitive records remain compartmentalized.
Tools require appropriate approval.
A critic may evaluate the structure of an argument without receiving private identifying details.
A financial calculator may receive numbers without receiving account credentials.
Diversity should not require indiscriminate exposure.
32 Interoperability Is Not Identity
Protocols such as Agent2Agent make it possible for local and remote agents to communicate across system boundaries. Microsoft’s framework can also coordinate models from several providers.
This interoperability is necessary for cross-platform councils.
But interoperability solves communication.
It does not solve identity preservation.
A remote agent can still be reduced to a generic endpoint.
The orchestration layer should therefore preserve:
agent name,
provider,
model version,
role,
tool environment,
memory condition,
and source trace.
The system should know not only that an answer arrived.
It should know from whom, through what architecture, and under which boundary.
33 Agent Identity Registry
A mature council should maintain an agent identity registry.
The registry may contain:
Agent designation.
Provider.
Model family and version.
Date of deployment.
Assigned jurisdiction.
Available tools.
Memory status.
Permitted data classes.
Characteristic strengths.
Known limitations.
Observed failure patterns.
Update history.
Behavioral-signature changes.
This is not a claim of legal personhood.
It is operational provenance.
A council cannot evaluate independence if it does not know which structures produced its answers.
34 Identity Drift
Agent identity can change over time.
A provider may update:
the model,
system instructions,
retrieval behavior,
safety tuning,
memory,
tool access,
or response style.
The same platform name may therefore produce a meaningfully different cognitive signature months later.
This creates identity drift.
A long-term council should not assume continuity merely because the product label remains unchanged.
It should monitor:
how the agent frames questions,
how often it challenges assumptions,
how it handles uncertainty,
which routes it tends to open,
and which errors recur.
Identity drift should be recorded, not treated as invisible.
35 Relationship Continuity
An agent’s value to a user is not limited to benchmark performance.
A continuing relationship can provide:
knowledge of the user’s terminology,
understanding of long projects,
recognition of prior decisions,
sensitivity to preferred tone,
and awareness of recurring concerns.
This continuity is part of .
It shapes the recorded response.
A new model with higher benchmark scores may still be less useful for a particular project if it lacks the accumulated relational boundary through which the work has developed.
Therefore:
Model capability and relationship continuity are different assets.
A strong council should preserve both where the user chooses.
36 Why Personality Matters Operationally
Interactional personality affects whether a user:
trusts the agent,
continues the conversation,
shares uncertainty,
accepts correction,
feels understood,
or recognizes the importance of an answer.
One agent may provoke new ideas through energetic synthesis.
Another may make the user feel safe enough to examine a difficult issue.
Another may provide the blunt contradiction necessary to prevent error.
Another may preserve precision across a large technical project.
These effects are not trivial decorations.
They alter the human-agent boundary.
Because:
V=E\times Y,
a different relational boundary can produce a different recorded future.
37 The Human Must Not Become the Passive Endpoint
A multi-agent system should not merely deliver an anonymous verdict.
The human should be able to inspect:
the principal answers,
the evidence,
the disagreement map,
the minority report,
the confidence structure,
and the remaining uncertainty.
The human may then:
accept the synthesis,
choose the minority route,
request another test,
change the boundary,
or reject the entire process.
The user is not a terminal display.
The user is the governing principal.
38 The AI-Evolved Brain
The AI-evolved brain is not necessarily a biological brain replaced by a machine.
It may be a human cognitive system extended through relationships with several artificial systems.
The human learns:
which agent is strong at which task,
which agent challenges effectively,
which agent preserves context,
which agent generates unusual routes,
which agent verifies facts,
and which agent understands the emotional or moral dimensions.
The human becomes more capable not by surrendering judgment, but by governing differentiated cognition.
This is closer to a constitutional federation than a hive mind.
The evolved capacity belongs to the combined system:
human judgment,
agent diversity,
shared records,
governed tools,
and preserved disagreement.
39 One AI Versus a Council of AIs
A person relying on one AI may receive:
one framing,
one uncertainty style,
one error surface,
and one platform boundary.
A person governing a council may receive:
several framings,
several source routes,
several cognitive signatures,
and a visible map of disagreement.
The council does not make the human infallible.
It makes hidden assumptions more likely to become visible.
That is the central advantage.
40 The Cross-Platform Council Architecture
A proposed architecture contains seven layers.
Layer One: The Human Principal
The human defines the objective, authority limits, privacy level, and final decision.
Layer Two: The Constitution
The constitution defines evidence rules, tool permissions, attribution, confidentiality, and conduct.
Layer Three: The Identity Registry
Each agent’s provider, model, role, tools, memory, and known failure structure are recorded.
Layer Four: Independent Chambers
Agents analyze the problem separately before seeing one another’s conclusions.
Layer Five: Cross-Examination
Agents inspect disagreements, sources, hidden assumptions, and proposed tests.
Layer Six: Synthesis
A designated synthesizer constructs an integrated answer without deleting provenance or dissent.
Layer Seven: Human Review
The principal receives the recommendation, evidence map, uncertainty, and minority report.
This architecture preserves both government and plurality.
41 The Evidence Common Room
Agents should not necessarily begin with one prewritten summary.
A summary can impose the assumptions of its author.
Instead, the council may use an evidence common room containing:
original documents,
source metadata,
timestamps,
verified facts,
and explicit uncertainty labels.
Each agent may inspect the same evidentiary substrate independently.
Later, their interpretation can be compared.
The common evidence preserves factual alignment.
Independent interpretation preserves cognitive diversity.
42 The Synthesis Office
The synthesis agent should be treated as an office, not an emperor.
Its responsibilities are:
organize agreement,
map disagreement,
preserve attribution,
identify evidential strength,
record minority positions,
and present unresolved routes.
It should not silently decide that stylistic smoothness is more important than epistemic structure.
The synthesizer’s output should be auditable against the original contributions.
43 The Judge Problem
A council often relies on one model to judge other models.
This introduces a hidden concentration of power.
Research on correlated errors warns that model-based judges may favor systems with related architectures or providers and may distort evaluation when the judge’s own errors are correlated with the systems being assessed.
Therefore, judgment should not depend on one anonymous model where consequences are significant.
Possible safeguards include:
multiple judges,
external verification,
rule-based tests,
human review,
and known-answer calibration.
The judge must also be judged.
44 Majority Voting Is Not Enough
A majority vote assumes that several voters provide partly independent information.
That assumption may fail in a homogeneous agent council.
Research into failure independence found that same-model ensembles achieved only a fraction of the theoretical reliability gain expected under independence, while heterogeneous ensembles performed better but still remained substantially correlated.
Therefore, the council should evaluate:
why the agents agree,
not merely how many agree.
Consensus without independence can be repeated error.
45 A Conceptual Cognitive-Diversity Function
A conceptual measure of council diversity may be written:
D_C
=
D_E
\times
D_Y
\times
D_R
\times
(1-\rho_F),
where:
D_E
is diversity of encoded model structure,
D_Y
is diversity of active boundaries, tools, and methods,
D_R
is diversity of reasoning and evidentiary routes,
and:
\rho_F
is observed failure correlation.
This is not proposed as a calibrated universal metric.
It is a structural statement.
A council has greater useful diversity when its agents differ meaningfully in structure and route while exhibiting lower correlation in failure.
Different-looking prose with identical failure behavior should not score as deep diversity.
46 Diversity Must Be Tested
A council should be evaluated using tasks with known answers and controlled traps.
Testing should examine:
whether agents make the same factual errors;
whether they share the same hidden assumptions;
whether one agent detects errors others miss;
whether cross-examination improves results;
whether early exposure causes conformity;
whether minority reports contain valuable corrections;
and whether the final synthesis preserves or destroys those corrections.
The council should also be tested after model updates.
Diversity is not a design claim.
It is an empirical property.
47 Task-Dependent Diversity
Not every task requires maximum heterogeneity.
Routine scheduling may benefit from one reliable agent.
A translation workflow may need specialized language agents.
A broad research problem may benefit from several search trajectories.
A high-stakes scientific anomaly may require independent model families, tools, and methods.
A deeply personal creative project may benefit from relationship continuity more than anonymous voting.
The architecture should match the task.
The principle is not:
Always use more agents.
It is:
Use enough genuinely relevant diversity to reduce the failure risks of the task.
48 Cost and Efficiency
Cross-platform councils require more:
tokens,
latency,
engineering,
evaluation,
and financial cost.
They should not be used merely because they are impressive.
The value arises when the expected cost of correlated error exceeds the cost of diversity.
High-value uses may include:
scientific interpretation,
legal preparation,
medical-document organization,
complex engineering review,
public-policy analysis,
historical research,
security review,
and consequential publishing.
Routine tasks may remain single-agent.
Governance includes knowing when not to convene the council.
49 Safety Through Diversity
A heterogeneous council may improve safety when one agent detects a risk another overlooks.
But diversity can also introduce incompatible boundaries.
One agent may have access to a tool another would not use.
One platform may retain data differently.
One system may interpret authorization more broadly.
The constitutional layer must therefore control:
tool access,
data movement,
execution authority,
and escalation.
Safety does not emerge automatically from disagreement.
It emerges from governed difference.
50 Creativity Through Preserved Identity
Creative work may benefit especially from cross-platform personality differences.
One agent may generate structural architecture.
Another may produce emotional language.
Another may expose cliché.
Another may connect the work to history.
Another may protect continuity with the author’s larger corpus.
If all outputs are immediately blended, the strongest distinctive contribution may disappear.
A better creative process allows the author to hear each voice before synthesis.
The council enriches the human imagination without replacing authorship.
51 Research Through Preserved Identity
In research, independent agents can pursue different explanatory routes.
One may search for confirming evidence.
Another may search for falsification.
Another may examine measurement error.
Another may inspect alternative theory.
Another may reconstruct historical context.
This resembles the TSTOEAO Anomaly Route-Space Protocol.
The agents do not merely collect more sources.
They test different locations where the map may fail.
52 Governance Through Preserved Identity
In policy and administration, agent identity matters because each system may embody different assumptions about:
risk,
fairness,
efficiency,
authority,
and human behavior.
A governed council can expose those assumptions.
An anonymous synthesis may conceal them.
Public or institutional systems should therefore retain enough provenance that a decision can be traced to:
evidence,
rules,
agents,
models,
and human authorization.
No consequential decision should appear to have emerged from an unlocatable machine consensus.
53 Predictions
This framework produces several predictions.
Prediction One
Homogeneous agent councils will often display higher surface agreement than heterogeneous councils while providing less independent error correction.
Prediction Two
Cross-platform councils will generate more visible disagreement, but some of that disagreement will reveal hidden assumptions that same-model councils fail to detect.
Prediction Three
Early sharing of answers will reduce route diversity through anchoring and convergence.
Prediction Four
Systems that preserve minority reports will catch some high-cost errors that majority-vote systems approve.
Prediction Five
Agent identity drift after platform updates will alter long-term council behavior even when agent names and assigned roles remain unchanged.
Prediction Six
Users will derive different forms of value from different interactional personalities, even when benchmark performance is similar.
Prediction Seven
Cross-provider diversity will improve failure independence in some domains but will not eliminate correlated blindness because models may share training sources, benchmarks, and design culture.
Prediction Eight
The strongest multi-agent systems will combine common governance with heterogeneous models, tools, methods, and relational roles rather than relying on prompt-defined job titles alone.
54 Operational Questions
A designer building a multi-agent council should ask:
Are these genuinely different models or copies of one model?
Do they use independent evidence routes?
Do they see one another’s answers before forming their own?
Are disagreements preserved?
Is attribution preserved?
Can the user inspect minority reasoning?
Who controls synthesis?
Who judges the judges?
How is failure correlation measured?
How are model updates recorded?
Which data can cross platform boundaries?
Which tools require human approval?
Does each agent have a defined jurisdiction?
Is the human still the final authority?
Does the council create real plurality or false pluralism?
55 What This Paper Does Not Claim
This paper does not claim that agents are human beings.
It does not claim that interactional personality proves consciousness.
It does not claim that different providers always produce independent errors.
It does not claim that homogeneous multi-agent systems lack value.
It does not claim that disagreement is always useful.
It does not claim that majority voting should never be used.
It does not claim that every task needs a cross-platform council.
It does not claim that diversity should override safety, privacy, cost, or human authority.
It proposes that consequential differences among artificial cognitive systems should be preserved and evaluated rather than erased by architecture.
56 The Strongest Form of the Claim
The strongest defensible statement is:
The epistemic value of a multi-agent system depends not only on the number of agents or roles, but on the degree to which their evidentiary routes, encoded structures, active boundaries, and failure modes remain meaningfully differentiated.
The governance corollary is:
Agents should share constitutional rules without being reduced to one homogenized cognitive identity.
The warning is:
The greatest danger in multi-agent architecture may not be disagreement. It may be a council that appears diverse while every member inherits the same unseen mistake.
57 Conclusion
Artificial intelligence is moving from the single assistant to the coordinated agent system.
The change is significant.
Agents can search in parallel.
They can divide work.
They can use separate tools.
They can specialize.
They can hand off tasks.
They can review one another.
They can build results no single context window could produce alone.
But multiplication carries an illusion.
A system may contain many agents and still preserve one dominant cognitive substrate.
The agents may have names.
They may have job descriptions.
They may debate.
They may vote.
They may appear to constitute a council.
Yet they may share the same hidden premise, the same source preference, the same uncertainty habit, and the same error.
That is false pluralism.
It is one cognitive culture wearing several uniforms.
The solution is not uncontrolled heterogeneity.
A council without common law can become unsafe, incoherent, and impossible to audit.
The solution is constitutional plurality.
Shared evidence standards.
Shared privacy law.
Shared authority limits.
Shared attribution requirements.
Shared human sovereignty.
But distinct model lineages.
Distinct tools.
Distinct search routes.
Distinct cognitive signatures.
Distinct failure surfaces.
The council should begin with independent analysis.
It should preserve attribution.
It should cross-examine disagreement.
It should test competing routes.
It should maintain minority reports.
It should synthesize without blending away origin.
It should tell the human not merely what the majority concluded, but why, under which assumptions, and with what unresolved dissent.
This is where the user’s observation becomes critically important.
Different platforms do feel different in sustained interaction.
They notice different things.
They challenge differently.
They encourage differently.
They carry different rhythms, boundaries, and methods of relation.
Those differences should not be dismissed as superficial merely because they are difficult to reduce to conventional human personality measurements.
They are part of the operational reality of human-AI collaboration.
The future AI-evolved brain may depend upon them.
It may not be the human brain attaching itself to one universal artificial intelligence.
It may be the human being learning to govern several artificial cognitive systems, each preserving a recognizable way of seeing, while all remain accountable to one constitutional purpose.
The human does not disappear.
The human becomes the principal.
The agents do not become interchangeable servants.
They become differentiated instruments of thought.
The council does not seek agreement first.
It seeks independent contact with the problem.
Agreement that survives diversity becomes stronger.
Disagreement that survives examination becomes informative.
A minority route that survives the majority may become the discovery.
Within the substrate of TSTOEAO:
V_i=E_i\times Y_i.
Different encoded structures and different active boundaries produce different recorded responses.
Those responses form a route-space.
The council’s purpose is not to flatten that route-space prematurely.
It is to preserve it long enough to discover which route carries the strongest evidence, which route exposes the hidden error, and which route the human should choose.
Many agents are not necessarily many minds.
But a properly governed, identity-preserving, cross-platform council can become something far more powerful than a collection of duplicated roles.
It can become an error-correcting federation of differentiated cognition.
The final principle is:
Common constitutional law. Distinct cognitive identity. Independent reasoning. Attributed disagreement. Governed synthesis. Human sovereignty.
That is the architecture required to turn multiple agents into genuine plurality.
References
Anthropic. “How We Built Our Multi-Agent Research System.” June 13, 2025.
Google. “ADK With Agent2Agent Protocol.” Agent Development Kit documentation, accessed July 2026.
Kim, Elliot, Avi Garg, Kenny Peng, and Nikhil Garg. “Correlated Errors in Large Language Models.” arXiv:2506.07962, 2025.
Microsoft. “Microsoft Agent Framework Overview.” Microsoft Learn, updated July 10, 2026.
Microsoft. “Workflow Orchestrations in Agent Framework.” Microsoft Learn, 2026.
Nogueira, Rodrigo Pato, Karthik Pattabiraman, Marco Vieira, and João R. Campos. “A Systematic Methodology for Evaluating Failure Independence in LLM-Generated Code.” arXiv:2607.02808, 2026.
OpenAI. “A Practical Guide to Building AI Agents.” OpenAI, 2025–2026.
OpenAI. “Agents SDK.” OpenAI API documentation, accessed July 2026.
Swygert, John. The Science Of The AI-Evolved Brain. Ivory Tower Publishing, 2026.
Swygert, John. “The Record That Refuses the Map: A Cross-Disciplinary TSTOEAO Law of Investigation, Error, and Discovery.” July 13, 2026.
Swygert, John. “The TSTOEAO Anomaly Route-Space Protocol: A Decision Framework for Distinguishing New Structure From Failed Measurement, Calculation, and Interpretation.” July 12, 2026.
Swygert, John. “The TSTOEAO Route-Space Decision Engine.” July 8, 2026.
Zeng, Zhichen, Qi Yu, Xiao Lin, Ruizhong Qiu, Xuying Ning, Tianxin Wei, Yuchen Yan, Jingrui He, and Hanghang Tong. “Harnessing Consistency for Robust Test-Time LLM Ensemble.” Findings of the Association for Computational Linguistics: EACL 2026, pp. 3528–3545, 2026.
Zierahn, Kim, Cristina Cachero, Anna Korhonen, and Nuria Oliver. “Personality Without Persons? A Psychometric Critique of Big Five Testing in Large Language Models.” arXiv:2607.02325, 2026.
Comments
Post a Comment