Architectural Insights

Practitioner guides on enterprise and agentic systems architecture — written from inside real architecture decisions, not vendor marketing.

Series

Foundations of Token-Efficient Context Design

Five parts on designing agentic systems that assemble context deliberately for each operation, instead of accumulating it until the context window is full. Starts with the architectural case for token efficiency, then works through minimal-and-complete context, authoritative state separation, normalized entities, and exact-wording preservation.

Part 1 of 5

Architectural Principles for Token-Efficient Agentic Systems

Why reliable agentic systems require context to be deliberately assembled for each operation rather than accumulated until the model context window is full.

Part 2 of 5

Minimal and Complete Model Context Design

A diagnostic guide to distinguishing incomplete agent context from inefficient context and selecting the right architectural response.

Part 3 of 5

Authoritative State Must Live Outside the Model Context

An architectural guide to separating durable, canonical system state from the transient context assembled for each model call.

Part 4 of 5

Normalized Entities and Context Slices Beyond Chunk Retrieval

Why agentic systems need normalized entities and operation-specific context slices when document chunks no longer provide reliable or efficient retrieval.

Part 5 of 5

Preserving Exact Source Wording in LLM Systems

Why retrieved facts, figures, quotes, and defined terms require architectural preservation controls rather than prompt instructions alone.