Why AI needs stricter architecture, not looser architecture
AI-assisted generation and agentic coding can accelerate delivery while dissolving system boundaries. Model calls, tool calls, persistence, prompts, policy checks, and domain behavior often collapse into one script, controller, or loop. MiRuntime calls this architectural liquefaction: infrastructure and authority spread until the system has no reliable place where business truth, permissions, or lifecycle rules are enforced.
- A raw model call inside business logic is not architecture.
- Prompt instructions are not authorization.
- Chat history is not a lifecycle state machine.
- Logs without typed events are not evidence.
- Tool descriptions are not policy.
The dependency rule remains useful because it keeps business rules independent from frameworks, databases, and provider SDKs. In MIR, that separation also prevents probabilistic output from becoming ambient authority. Source: Microsoft architecture guidance Source: Prompt-injection guidance
Clean Architecture layer mapping
| Clean Architecture Layer | MIR Responsibility | Must Not Depend On |
|---|---|---|
| Domain / Entities | Run, objective, risk class, approval, evidence event, tool action, memory scope, model route | LLM providers, databases, web frameworks |
| Use Cases / Application | Runtime workflows, validation, policy checks, recovery, approval orchestration | Provider-specific APIs, direct SQL, prompt-only logic |
| Interface Adapters | Model adapters, tool connectors, repositories, controllers, presenters | Inner-layer implementation details |
| Frameworks / Drivers | WordPress page, API host, queue, model server, database, object store, observability stack | Runtime business rules |
Logical dependency direction matters more than deployment shape. A modular monolith can follow this rule; a network of microservices can violate it.
Ports and adapters
A runtime use case should depend on interfaces defined toward the core. Concrete adapters translate those interfaces into provider SDKs, protocols, queues, stores, and observability backends.
IModelAdapterIToolConnectorIPolicyEngineIEvidenceLedgerIMemoryProjectionStoreIRuntimeStateStoreIApprovalGatewayICheckpointStore
The use case therefore does not depend on OpenAI, Anthropic, Gemini, vLLM, Triton, ONNX Runtime, PostgreSQL, Redis, Kafka, S3, or a specific cloud. Those technologies can be valid adapters; they are not the runtime’s domain model.
Deterministic boundary
What belongs outside the prompt
Prompt text can describe intent to a model. It must not be the only enforcement point for any of these controls.
Capability-first organization for AI-assisted development
Traditional layer-first repositories can scatter one capability across many distant directories. Human developers may internalize that map, while coding agents repeatedly spend context on navigation and unrelated abstractions. A capability-first layout keeps a feature’s domain types, use cases, ports, adapters, tests, and intent documents close enough to load together without reversing the dependency rule.
/src
/Runs
/Domain
/UseCases
/Ports
/Adapters
/Tests
AGENTS.md
/Policies
/Domain
/UseCases
/Ports
/Adapters
/Tests
AGENTS.md
/Evidence
/Domain
/UseCases
/Ports
/Adapters
/Tests
AGENTS.mdA root-level AGENTS.md, CLAUDE.md, or equivalent intent file should describe the architectural constitution: dependency direction, invariants, forbidden shortcuts, source-of-truth boundaries, required tests, and the controls that may never be delegated to prompt text. Capability-level files can narrow that context without weakening the root rules.
Ports, adapters, dependency inversion, and explicit authorization are production practices. Capability-first organization for coding-agent context is an emerging practice and should be evaluated against repository size, team ownership, and tooling.
Source record
References
Microsoft. Microsoft Learn. Published Current documentation; last reviewed 2026-06-23 UTC. Official architecture guidance.
National Institute of Standards and Technology. NIST. Published 2023-01-26; last reviewed 2026-06-20 UTC. Government framework.
- LLM01:2025 Prompt Injection Primary source
OWASP GenAI Security Project. OWASP. Published 2025; last reviewed 2026-06-20 UTC. Security guidance.