CHAI describes a topology, not a product
CHAI treats an AI system as a set of capabilities that make and combine judgments.
This differs from a multi-agent arrangement that assigns several LLMs different prompts and roles. That pattern can be useful. It still shares a large part of its failure surface when every agent depends on the same model family, the same provider, the same context assumptions, or the same orchestration layer. Recent research on multi-agent LLM failures similarly finds that many failures originate in system design rather than simple prompt-following defects. Why Do Multi-Agent LLM Systems Fail? CHAI is useful when the task itself calls for different kinds of computation and when the system must retain a legible account of how their outputs became an action.
Research on composable AI makes the same basic distinction.
Heterogeneity changes the engineering problem
Components with different jobs also have different operating characteristics. Treating those outputs as interchangeable makes the coordination layer unreliable before it has made a single bad prediction.
Consider a supply-chain monitoring system. A coordinator receives four different kinds of evidence. Its job is not simply to average them.
It must know what each output means, how recently it was produced, what conditions invalidate it, and what decision it is permitted to influence.
That distinction also prevents a common category error: using a language model where a structured model, a deterministic rule, or an optimizer is the better fit.
The question is not whether one substrate is more advanced. The question is what evidence the task presents and what error the organization can tolerate.
Recursion makes governance cumulative
A CHAI system becomes a hive when an aggregation node can itself become an input to a higher-order decision.
A flat architecture can often get away with one coordinator. A recursive architecture cannot. It needs a decision contract at every aggregation point. Four elements belong in that contract.
Input semantics
The parent needs more than a value. It needs the output type, provenance, freshness, validity range, and any confidence or uncertainty measure the child can support. A score from one model does not automatically mean the same thing as a score from another. Research on neural-network calibration makes the narrower point: a model's reported confidence need not represent its empirical likelihood of being correct. Guo et al., On Calibration of Modern Neural Networks Normalizing or comparing signals across components therefore requires an explicit method, not a convenient assumption.
Degraded-mode behavior
A parent must know what to do when a child times out, returns an invalid result, falls below an evaluation threshold, or is operating outside its approved conditions. It may route to a fallback, lower the scope of the decision, hold the case for review, or stop the workflow. The correct choice depends on the decision. Leaving the choice implicit turns an integration failure into an untraceable business decision. The broader distributed-systems discipline makes the same point: circuit breakers isolate a faulting dependency and permit graceful degradation rather than letting it overload the wider workload. Microsoft Azure Architecture Center
Arbitration
Two children can produce conflicting outputs without either being defective. A fraud score may conflict with an identity check. An optimizer may recommend an action that a policy engine forbids. The architecture must state whether a rule can veto, whether one source receives priority for a defined condition, whether a human reviews the case, or whether the system declines to act. Arbitration is an authority decision, not a weighting detail.
Traceability
The system must preserve the inputs, versions, routing decisions, and final outcome needed to reconstruct a material decision. NIST's AI Risk Management Framework calls for production monitoring of AI systems and their components, documented limitations, and post-deployment mechanisms for override, incident response, recovery, and change management. Those are operating requirements, not observability decoration. NIST AI RMF 1.0 W3C PROV provides a mature, technology-neutral provenance model for recording the entities, activities, and responsible agents that produced an artifact. W3C PROV-DM
These four elements are the substantive core of CHAI. A diagram of connected models does not establish any of them.
Interfaces make replacement possible
The useful test of modularity is not whether a vendor can add another feature. It is whether the organization can replace a component without renegotiating every downstream dependency.
That requires an interface contract that names the following.
It also requires portable evaluation data. If a replacement component cannot be tested against the cases that justified the incumbent, the organization cannot distinguish a safe swap from a cosmetic one. NIST's Generative AI Profile identifies the accountability problem directly: systems composed from third-party models, data, and software can make it difficult to attribute behavior to any one source. It recommends continuous monitoring of third-party systems and documented fallback arrangements. NIST AI 600-1
Modular Open Systems Approach
MOSA requires modular interfaces, verification against appropriate standards, machine-readable interface definitions, and architectures that permit components to be added, removed, or replaced over the life of a system. DoD MOSA
The Department of Defense's approach offers a useful analogy, not a proof that an AI architecture is sound.
The governance questions
CHAI applies the same discipline to AI decision components, then adds the governance questions that emerge when those components contribute judgments rather than merely exchange data.
Open standards alone do not settle confidence, authority, or fallback behavior. They establish the seam. CHAI specifies what must happen across it.
The cost belongs in the architecture decision
A hive architecture earns its cost only when the decision problem needs it. Every additional node adds interface maintenance, evaluation work, monitoring, coordination logic, and potential latency. A simple task with one dependable model and a clear human checkpoint does not become better because it has been decomposed into a network.
Contract ownership
Someone must maintain each boundary as components evolve.
Observability
A production team needs to trace a result through the decisions that formed it, including the versions and degraded conditions that were present at the time.
Coordination
Aggregation and arbitration take time, and latency budgets should constrain topology for time-sensitive paths.
A safety-critical workflow, a regulated decision, or a multi-modal process with independently changing components.
A narrow classification task.
These costs are not reasons to collapse back to a monolith. They are reasons to assign architecture only where the system's decision structure warrants it.
