Brief #137

Graph-based reranking improves retrieval AND reduces harmful distractors, while agentic workflows amplify context failures when models generate their own noise. Context quality compounds or degrades depending on retrieval method choice.

Agent memory requires four distinct storage types (episodic/semantic/procedural/long-term) that must be retrieved selectively and compiled into working context per-prompt. Architecture determines what gets pulled; not all context goes everywhere.

Multi-agent systems succeed with centralized orchestrators spawning parallel specialists constrained by tool access and shared artifacts. Flat hierarchies and redundant context rearrangement create 15x token burn without quality gain.

Models trained on billions of invocations with specific JSON structures develop implicit expectations about interface format. Switching harnesses fights learned patterns, preventing effectiveness from compounding.

Context requires systematic lifecycle management (Generate → Evaluate → Distribute → Observe) with versioning, testing, and team practices. Ad-hoc context management is the equivalent of coding without version control.

Modern models infer process FROM outcome definition rather than FROM process instructions. Clarity of END STATE matters more than phrasing of steps.

For procedural tasks with known workflows, embedding the procedure in system prompt and allowing model self-orchestration produces better results than external state management frameworks.

Vendors market context window ceilings but obscure practical constraints: max output tokens on API tiers, rate limits, and pricing structures create different effective windows than advertised specs.