Brief #128

Protocol standardization (MCP) solves context connectivity but doesn't automatically preserve intelligence across sessions. You still need explicit state files, checkpoint patterns, and configuration management to prevent context reset.

Single-vector optimization on system prompts (reducing verbosity, constraining length) can simultaneously improve token efficiency while degrading reasoning quality on edge cases. Requires per-model evals across broad production contexts, not just targeted optimization.

Standard RAG benchmarks miss the real problem: retrieval systems produce ranked, semantically similar distractors that degrade LLM reasoning more than random noise does. Graph-based reranking recovers up to 44% performance by filtering harmful context.

Giving agents maximum action space with direct tool access + error visibility enables better self-correction than carefully wrapped abstractions. Wrappers encode assumptions that break when environments change; raw access lets agents adapt using their training knowledge.

Open-source 27B parameter models running locally on consumer hardware now match Claude Opus performance on real coding tasks. This shifts context engineering from API optimization to local state management—no latency, full context control, zero cost per call.

MCP protocol now has multiple versions (2025-03-26, 2025-06-18) with different OAuth requirements and structured output expectations. Context validation must explicitly handle version compatibility or suffer silent degradation.

Most multi-agent orchestrators resemble 'ant farms' (emergent behavior, no coordination) when they should be 'software factories' (explicit context handoffs, structured coordination). Builders lack mental models for agent coordination architecture, not better agents.