← Latest brief

Brief #151

50 articles analyzed

Context engineering is fracturing into two distinct disciplines: infrastructure-level protocol standardization (MCP ecosystem convergence) versus application-level context architecture (how practitioners actually structure information for agents). The gap between them is widening—vendors solve connectivity while practitioners rediscover that information design matters more than tooling.

◐

RAG Systems Fail on Fast-Moving Codebases

CONTRADICTS retrieval-augmented-generation — existing graph treats RAG as stable solution, practitioners abandoning it for velocity reasons

Semantic search and vector databases break when source-of-truth updates faster than index refresh cycles. Practitioners replacing static RAG with live codebase traversal and dynamic context assembly.

→ Replace static vector search with dynamic context assembly for repositories that change multiple times per day. Assign Agent Manager role to own context freshness.
@dani_avila7: At my previous startup we tried RAG with semantic search and vector DBs

Abandoned RAG+vector DB after index drift made agent outputs unreliable. Switched to live traversal with Agent Manager role owning CLAUDE.md context.

Prompt Engineering Best Practices 2026

Author realized context assembly failures (wrong retrieval, missing tool definitions) caused output drift despite correct prompts. Static indexes lag behind reality.

AI Coding Agents in 2026: Claude Code, Cursor, and How We Actually Use Them

2-person team maintaining 12 repos discovered CLAUDE.md context files and memory systems outperform retrieval for fast-changing codebases.

More signals

Context Structure Beats Prompt Engineering for Token Efficiency

EXTENDS prompt-engineering — existing graph treats prompting as primary technique, this establishes information architecture as separate prior concern

Converting prose to structured formats (tables, YAML, flow diagrams) delivers 10x token savings with better AI comprehension. Context engineering emerges as distinct discipline from prompt engineering.

→ Audit your prompts for prose descriptions. Convert to structured formats: use tables for comparisons, YAML for configurations, flow diagrams for processes. Measure token reduction.
Prompt engineering was 2023. RAG was 2024

Author empirically demonstrated same information in structured format (API→Queue→Workers→DB) consumes fewer tokens and delivers clearer AI understanding than prose.

●

MCP Adoption Creates New Authorization Context Layer

EXTENDS model-context-protocol — existing graph documents MCP as integration standard, this identifies security/permissions as next architectural layer

Model Context Protocol standardizes tool integration but exposes gap: agents need machine-readable permission context, not just capability discovery. Authorization becomes context engineering problem.

→ When implementing MCP servers, design permission schemas that agents can query programmatically. Treat authorization as discoverable context, not hidden backend logic.
AI Agents, the Model Context Protocol, and the Future of Authorization Guardrails

MCP solves 'what tools exist' but creates new problem: 'what is agent allowed to access?' Permissions must flow as context to agent, not just exist in backend.

Agents Need Constraint Architecture Not Better Reasoning

EXTENDS agent-design-patterns — existing graph documents orchestration, this identifies constraint architecture as foundational prerequisite

Production agent reliability comes from structuring the problem space (verifiable constraints, strategic walls, goal regions) rather than improving model capability. Constraint design is context engineering.

→ Stop investing in better prompts. Design verifiable success criteria, explicit boundaries, and role constraints. Architect the problem space before deploying agents.
@fchollet: A mental model for working with coding agents

François Chollet frames agents as 'blind squirrels'—success requires architecting problem space with verifiable constraints and strategic boundaries, not relying on agent reasoning.

◐

Statefulness Requires First-Class Architectural Design Not Bolt-On Memory

EXTENDS state-management — existing graph treats state as storage problem, practitioners discovering it's identity/architecture problem

Practitioners cannot retrofit statefulness onto stateless agents. State representation must be core design decision that shapes agent identity and reasoning across sessions.

→ If building stateful agents, design state representation architecture before writing any agent code. Define what gets preserved, how identity persists, and how state influences reasoning.
@yoheinakajima: maybe you can't just tack statefulness onto an agent

Yohei Nakajima observes that statefulness requires rethinking agent representation from first principles, not adding memory as feature.

LLM Delegation Without Reasoning Trace Creates Learning Black Holes

EXTENDS human-ai-collaboration — existing graph documents collaboration patterns, this identifies reasoning preservation as critical success factor

Using LLMs to skip difficult reasoning conversations eliminates team learning substrate. Intelligence compounds only when reasoning process is captured, not just outputs.

→ When using LLMs for team decisions, capture the reasoning dialogue not just the output. Use adversarial questioning patterns to force explicit reasoning traces.
@GergelyOrosz: Situation 1 vs Situation 2 delegation patterns

Orosz contrasts human-facilitated disagreement (learning preserved) versus LLM delegation (learning lost). Delegation without documentation creates knowledge dead-zones.

◐

KV Cache Prefix Reuse Enables Context Compounding at Scale

EXTENDS context-window-optimization — existing graph focuses on truncation strategies, this identifies caching as infrastructure-level solution

Serving-stack KV cache optimization with prefix caching allows long-context systems to preserve rich history without cost explosion. Infrastructure change unlocks intelligence compounding.

→ Audit your prompt structure for reusable prefixes (system instructions, static context). Insert explicit cache markers. Measure cost reduction from prefix reuse.
KV Cache Optimization for LLMs 2026: Engineering Guide

Prefix caching via explicit cache-control breakpoints in prompts enables fine-grained context reuse. Application-side lever for managing context costs at scale.