The Architecture of Intelligence

The system of intelligences may be the relevant unit of design

The first version began with an accessible allocation heuristic: put language models where language and interpretation matter; statistical models where accumulated data supports prediction; optimization where objectives can be formalized; deterministic controls where truth must not drift; and humans where authority, accountability or relationships matter. EVO retains that heuristic while surrendering any suggestion that human–machine complementarity, distributed cognition or joint social and technical design originates here.

The deeper question is not which component receives which task. It is what cognitive system must exist for an outcome to occur reliably. A consequential decision may move through historical data, a predictive model, deterministic eligibility constraints, language-model interpretation of ambiguity, human judgment, authorization, an audit record, outcome feedback and later institutional learning. The behavior belongs to the chain of representations and relationships, not only to one part.

EVO’s revised working thesis is therefore a synthesis and application claim. Architecture designs both components and the relationships among them: what each component can see, how representations change across boundaries, what can interrupt, which outputs constrain later action, where authority sits, how disagreement survives, and whether somebody can reconstruct the reasoning after the fact.

Cognition was already larger than the individual

Edwin Hutchins’s distributed-cognition work is foundational prior art. In Cognition in the Wild, ship navigation is analyzed as cognition distributed across people, artifacts, representations and coordinated activity. The implication for this essay is material: a component list—human here, model there—can miss the cognitive work performed by the relationships and transformations among them.

The sociotechnical tradition reaches further back. Trist and Bamforth’s 1951 study of coal-mining work examined technological change together with social organization rather than treating the technical system as independently optimizable. EVO did not originate joint design. The modern application is to resist optimizing an AI component while ignoring authority, workflow, incentives, interfaces, information structures and the people expected to absorb its failures.

Hybrid-intelligence research also explicitly defines complementarity between human and machine intelligence. Dellermann and colleagues describe systems intended to achieve better results by combining complementary capabilities. The claim that humans and AI can compensate for one another is therefore prior art. What remains open is whether general-purpose, generative and increasingly agentic systems alter the design problem enough to require new architectural methods.

A good component does not guarantee a good outcome

The intervention-ensemble concept in healthcare makes a strong version of this argument. An ML model does not improve patient outcomes by itself; benefit depends on the ensemble of knowledge, practices, conditions, monitoring and workflow in which it is embedded. This is not evidence that every company should borrow clinical governance wholesale. It is evidence against evaluating an intelligent component as though its measured performance were the system outcome.

This prior art strengthens an independent EVO observation while narrowing its novelty. Working software can solve the wrong problem. A technically functional system may optimize transaction mechanics while missing customer desire, institutional incentives or the work required around the tool. EVO’s contribution here is not discovering whole-system effects. It is applying that discipline to AI-native company architecture and treating commercial validity as a property of the intervention system, not the model demo.

Human oversight can fail while remaining visibly present

Research on automation misuse and disuse predates current generative AI. Parasuraman and Riley described misuse as overreliance that can produce failures of monitoring and decision bias. Contemporary evidence further complicates the easy promise of complementarity: a 2024 meta-analysis in Nature Human Behaviour found that human–AI combinations were, on average, worse than the better of humans or AI alone across included experiments, even though they often improved on humans alone.

Human authority at consequential boundaries is therefore not a magic safeguard and not an original EVO principle. A person can be technically ‘in the loop’ while lacking time, context, real discretion or the ability to detect machine error. The WHO’s AI-for-health guidance emphasizes autonomy, transparency, responsibility and accountability, but implementation remains the hard part. Meaningful authority requires a designed capacity to understand, disagree, halt, escalate and own consequences.

The strongest counterargument to EVO’s thesis is that deliberate composition may add enough interfaces, governance and representational translation to make the system less intelligent and less accountable. A simple expert with a checklist may outperform a heterogeneous architecture nobody understands. Additional intelligence can reduce system intelligence when it adds ambiguity faster than compensating capacity.

The artifacts are becoming participants

Traditional cognitive artifacts were often comparatively passive. Modern AI components interpret, generate, recommend, predict, communicate, adapt, initiate and increasingly act. The distinction is not absolute—older automated systems acted, and humans remain involved in contemporary systems—but the breadth and generality of machine participation may change the density of distributed cognition.

Responsibility may remain formally centralized while causal cognition becomes distributed. The accountable executive may not understand the complete chain by which a recommendation emerged. A model may influence deterministic eligibility, an agent may choose which evidence to retrieve, and a human may authorize an output framed by machine-generated context. Naming the system ‘intelligent’ can help analysis while also obscuring who had power to design, deploy, supervise and stop it.

This produces a limit question EVO does not yet know how to answer: how much orchestration complexity can an organization tolerate before its intelligence becomes illegible? Some boundaries should remain deliberately simple, deterministic and boring because predictability is more valuable than additional cognition. The ability to refuse intelligence may become an architectural capability.

What belongs to the literature—and what remains EVO synthesis

Established prior art: cognition can be distributed across people and artifacts; social and technical systems should be jointly considered; human–machine complementarity is a defined research program; real-world outcomes depend on intervention ensembles; automation can produce overreliance; and formal human oversight does not guarantee effective control.

EVO observation: capability is not permission; working software may solve the wrong commercial problem; conversation should not automatically become authoritative memory; and additional intelligent output can create noise rather than value. These observations are situated and do not establish general law.

EVO synthesis: apply distributed cognition, sociotechnical joint design, hybrid intelligence, intervention-ensemble thinking and authority design to the architecture of AI-native organizations. Working EVO hypothesis: the system of intelligences and their relationships may be the relevant unit of design, provided accountability and reasoning remain legible. This may turn out to be wrong.