Recognize the architecture before you rebuild the failure.
Use patterns and anti-patterns to identify whether an AI application keeps identity, context, policy, state, actions, and evidence connected.
Start with the application responsibility this page owns.
Connect the concept to evidence and implementation.
Eight questions to take into your next design review
The resource helps teams challenge an architecture before production exposes the same known failure modes.
Authority
Where does user, workload, tenant, purpose, and delegated authority survive?
Context
Which information is eligible, fresh, scoped, and provenance-backed?
Contracts
Where are schemas, side effects, and result meaning explicit?
Evidence
Can the team explain what happened after the fact?
Failure modes to recognize early
These patterns often look productive during prototypes and become expensive when the application has consequences.
LLM Everywhere
Using inference where deterministic computation or policy should own the step.
Prompt Spaghetti
Encoding application behavior in prompts without explicit contracts or versions.
Invisible Context
Supplying useful information without preserving eligibility, freshness, scope, and provenance.
Policy After Deployment
Reviewing governance only after the application path already exists.
Logging Instead of Replay
Collecting events without preserving the relationships needed to explain outcomes.
Autonomous Action Without Identity
Letting actions occur without explicit authority, operation identity, and effect evidence.
Reusable solution structures
A pattern is a reusable architecture structure, not a product feature checklist.
Deterministic-First Routing
Prefer deterministic computation when it can satisfy the contract before escalating to inference.
Confidence Escalation
Escalate uncertain work to stronger models, more context, verification, humans, or abstention.
Governance by Construction
Put authority, policy, obligations, and evidence into the application path.
Human Approval State
Bind review to the exact proposal, evidence, versions, and resumption semantics.
Permission-Aware Evidence
Make review possible without turning audit access into unrestricted data access.
Side-Effect Boundary
Separate proposed, attempted, ambiguous, confirmed, and failed external effects.
Use patterns before the failure is expensive.
Bring one proposed AI application into an architecture review and test it against the recurring patterns and anti-patterns.