← Latest brief

Brief #191

18 articles analyzed

Context engineering is inverting: practitioners are abandoning prompt optimization for state management infrastructure—30% of engineering time now spent on harness/memory/skills systems that preserve intelligence across sessions. The bottleneck has moved from 'better prompts' to 'persistent context architectures.'

◐

State Files Over Prompts: Loop Architecture Displacing Prompt Engineering

EXTENDS state-management — confirms state as critical but adds new dimension: state files as primary interface replacing prompts

Practitioners are replacing prompt optimization with explicit state management systems (text-based state files, loop architectures) that preserve agent context across sessions. The shift: intelligence compounds through persistent state rather than resetting with better instructions.

→ Allocate 20-30% of agent development time to building state persistence infrastructure (state files, harness systems, memory layers) rather than prompt tuning. Architect loops with explicit success criteria and hard boundaries.
@ianlapham: I probably spend 30% of my time as an engineer on harness, skills and memory

Practitioner reports 30% time allocation to context/state systems, predicts shift to 90%. Direct testimony that harness/memory work dominates over traditional engineering.

@EXM7777: i'm going to teach you how to run Fable 5 on autopilot, using my own library

Loop architecture using state files where 'model reads its own history before every round' to prevent redoing work. Explicit compounding mechanism through readable state persistence.

@ClaudeDevs: There's a lot of talk right now about 'designing loops' instead of prompting

Anthropic explicitly positioning 'loop design' as successor to prompting. Three required elements: success criteria, loop boundaries, external state awareness.

More signals
●

Multi-Tier Model Orchestration: 92% Performance at 37% Cost

EXTENDS multi-agent-orchestration — adds specific cost/performance trade-off data and advisor/executor pattern to existing orchestration concepts

Advisor/executor splits recover 92% of premium model performance using mid-tier executors with strategic advisor calls. The pattern: concentrate expensive reasoning at coordination layer, distribute cheap execution at scale.

→ Redesign single-model agents as two-tier systems: expensive model handles strategic decisions/guidance (called once per task), cheap model executes with that context. Target 90%+ performance at 60-70% cost.
@shao__meng: 方向：Executor → Advisor（执行者主动求助）

Practitioner testing pull (executor asks advisor) vs push (orchestrator delegates) models. Reports cost optimization while maintaining quality through strategic model role assignment.

Context Routing as Infrastructure: Algorithmic Not LLM-Based

EXTENDS context-window-management — adds algorithmic routing layer above existing management techniques

Production systems route context decisions algorithmically rather than delegating to LLMs. Runtime context strategy selection (write state, compress, retrieve, isolate subagent) reduces costs and latency by avoiding expensive LLM routing calls.

→ Build context routers that make deterministic decisions (if/else, retrieval, compression rules) before invoking LLMs. Reserve model calls for ambiguous reasoning, not context selection.
Context Orchestration: What It Is & How It Works - Redis

Explicit recommendation: determine context strategy algorithmically at runtime based on context type/triggers. Reserve LLM calls for nondeterministic tasks, not routing.

◐

MCP as Context Protocol: Standardizing Service Definition Context

EXTENDS model-context-protocol — confirms MCP importance but adds clarity on its role as context formalization layer

MCP adoption reflects shift from embedding context in prompts to formalizing it as machine-readable protocol. Service definitions (purpose, inputs, outputs, capabilities) become explicit contract layer rather than inferred from natural language.

→ Migrate from prompt-based service descriptions to MCP servers. Define service purpose, input/output schemas, and capabilities as structured protocol rather than natural language instructions.
Building with MCP and the Claude API - YouTube

MCP video articulates systematic approach to connecting agents to external systems while preserving context about tool capabilities across multi-turn interactions.

Tool Definition Pressure: RAG for Tool Discovery

EXTENDS tool-integration-patterns — adds specific solution (RAG) to known problem (tool scaling)

Multi-tool agents exhaust context windows on tool definitions before processing queries. Solution: treat tool catalogs as RAG corpus, semantically retrieve relevant tools per query rather than loading entire catalog.

→ For agents with 10+ tools: implement semantic search over tool definitions, inject only top-k matched tools per query. Monitor context consumption from tool definitions separately from task context.
Context Engineering for LLMs: Strategies and Patterns - n8n Blog

Explicit problem statement: tool definitions create context pressure. RAG-based tool discovery enables selective loading of only relevant tools per query.

Enterprise Context Governance: Judgment Requires Institutional Knowledge

EXTENDS retrieval-augmented-generation — institutional context governance is specialized form of RAG for enterprise settings

Autonomous agents fail in enterprise not from model limits but from missing institutional context (margins, strategy, permissions, definitions, lineage). Context governance is prerequisite for agent judgment, not a nice-to-have.

→ Audit what institutional knowledge your agents need (definitions, constraints, priorities, historical patterns). Build structured knowledge bases or governance layers BEFORE deploying autonomous agents.
For Enterprise AI, It's Not The LLM, It's The Context

Enterprise AI failing because models lack institutional context to judge outputs. Problem isn't capability but context governance (seasonal patterns, margins, strategy, permissions).

Integration Context Over New Tools: AI Sandwich Adoption Pattern

High-impact AI workflows integrate into existing information flows (email, calendar, Slack) rather than introducing new tools. The 'AI sandwich' (human intent → agent execution → human judgment) preserves context at handoff boundaries.

→ Before building custom AI tools, identify existing information flows (email threads, Slack channels, calendar patterns). Design agents to augment those contexts rather than replace them. Preserve human decision-making at input/output boundaries.
@trueventures: The AI workflows with the biggest impact usually aren't the fanciest

Impact workflows integrate into existing contexts (calendar/inbox) not new tools. AI sandwich model preserves human intent and judgment at boundaries. Failures become training data in feedback loop.