AI systems
This page defines system archetypes built on top of the infrastructure layer: retrieval systems, workflow agents, and decision-support surfaces. Cursor agents should treat these as templates when proposing new AI-powered capabilities.
Retrieval systems
Retrieval systems connect enterprise knowledge sources to deterministic model behavior. They are defined by their indexing strategy, retrieval logic, and evaluation harness.
- Ingestion: controlled connectors and chunking strategies for documents, tickets, and logs.
- Indexing: embedding, sparse, or hybrid indexing with versioned configurations.
- Retrieval: ranking logic, filters, and safety constraints applied before model invocation.
Retrieval constraints
- All retrieval behavior must be observable and reproducible based on logged configuration and queries.
- Changes to indexing or retrieval parameters are tracked and can be linked to evaluation results.
- Prompt construction uses retrieved context in a controlled, auditable way.
Workflow agents
Workflow agents are deterministic orchestrations of tools, retrieval steps, and model calls. They operate within clear boundaries rather than acting as unconstrained general agents.
- Tooling: explicit tool interfaces with arguments, preconditions, and failure handling.
- State: scoped state machines or workflows rather than unbounded conversation logs.
- Controls: policy-enforced actions, human-in-the-loop steps, and reversible operations.
Agent guarantees
- Each workflow agent has a defined domain, inputs, and outputs.
- All tool calls and decisions are logged and can be replayed for audit.
- Policy engines can intercept or veto actions that violate compliance constraints.
Decision-support surfaces
Decision-support systems present model outputs as structured recommendations, not opaque answers. They are designed to preserve traceability from input to decision.
- Inputs: clearly defined signals and context sources.
- Outputs: recommendations with confidence, justification, and supporting evidence links.
- Feedback: structured feedback mechanisms that feed into evaluation and retraining loops.