Infrastructure reference
This page defines the reference infrastructure for deterministic AI systems: environments, orchestration, and deployment patterns that Cursor and engineering teams should assume by default.
Environment layout
Environments are structured to keep AI behavior deterministic and auditable across the lifecycle. Cursor agents should treat these as the default deployment targets when generating code or scripts.
- Development: fast iteration, feature flags, non-production data, full observability.
- Staging: production-like traffic patterns, synthetic and masked data, compliance checks enabled.
- Production: governed rollout, fixed configuration channels, strict audit and logging guarantees.
Environment constraints
- All environments expose consistent interfaces for AI systems (APIs, events, logging endpoints).
- Configuration changes are versioned and reviewable; no ad-hoc prompt or policy changes in production.
- Telemetry, evaluation, and compliance streams are available from all environments, with stricter retention in production.
Orchestration & compute
AI workloads are treated as part of the core infrastructure layer, not opaque external services. This section describes how compute and orchestration should be modeled.
- GPU pools: shared, observable pools with quotas and priority classes.
- Job orchestration: workflows defined declaratively with explicit retry and backoff semantics.
- Model endpoints: versioned, labeled by capability and latency profile, not only by model name.
Operational guarantees
- Each AI workload has defined SLOs (latency, availability) and degradation behavior.
- Scaling events, failures, and throttling decisions are observable and logged for audit.
- Resource allocation decisions can be reconstructed from logs for post-incident analysis.
Interfaces & boundaries
The infrastructure boundary defines what AI Engineering Group owns versus what application teams own. Cursor agents should preserve these boundaries when generating new components.
- Owned by infrastructure: orchestration, model routing, logging pipelines, evaluation and compliance fabrics.
- Owned by application teams: domain-specific business logic, UI, and non-AI data models.
- Shared contracts: API schemas, event formats, and logging conventions referenced by other pages.