The Field Guide
Canonical definitions and how-tos for agentic execution systems. Direct answers first, evidence throughout, one meaning per term.
Foundations
The Agentic Execution Stack
The seven-layer framework for reliable AI agents: identity, skills, loops, tools, memory, evaluation, governance.
The Loop Reliability Model
Five axes for judging any agent loop: completion rate, verification quality, recoverability, observability, boundedness.
The Memory Safety Model
Six axes for judging agent memory: relevance, freshness, privacy, retrievability, editability, failure handling.
The Agentic Maturity Model
Six levels from ad-hoc prompting to governed execution systems. Find your level, and what to build next.
Governance for AI Agents: The Working Guide
Permissions, gates, logs, and accountability: the layer that turns an agent pilot into a deployment that passes audit.
Instruction files
The Complete Guide to SKILL.md
What SKILL.md is, its anatomy, how triggering works, and what separates a high-leverage skill from a renamed prompt.
What Is AGENTS.md?
The repository-root instruction file that tells any AI agent how to work in your project. A README written for agents.
What Is CLAUDE.md?
Claude Code's project instruction file: what it is, what belongs in it, and how to keep it from bloating.
SKILL.md vs AGENTS.md vs CLAUDE.md
The three file formats compared: procedure versus identity, when each applies, and how they work together.
How to Write a High-Leverage SKILL.md
Six steps from blank file to a skill that fires reliably, verifies its own work, and survives review.
Cursor Rules: Best Practices
Scoping, rationale, and the .mdc format: how to write Cursor rules that fire when they should and age well.
Loops, memory, and evaluation
Agent Loops, Explained
What agent loops are, why agents stop early or declare false success without them, and the named patterns that fix it.
Memory for AI Agents: The Field Guide
The four memory types, the storage and retrieval trade-offs, and why staleness handling decides whether memory helps or harms.
Evaluating Agent Skills
A practical rubric for vetting third-party skills before adoption: triggers, completeness, boundaries, evidence, docs.
What Is the Model Context Protocol (MCP)?
The open protocol that standardized the agent tool layer: how it works, why it won, and how to adopt servers safely.
Securing Agent Systems: Injection, Poisoning, and the Skill Supply Chain
The three attack surfaces every agent deployment carries, and the layered defenses that actually work.
Subagents and Orchestration, Explained
When to fan out, when to hand off, and why fresh context is the most underrated property of multi-agent design.
Choosing an Agent Orchestration Framework
LangGraph, CrewAI, the vendor SDKs, or none: a decision guide ranked by verification rigor, not demo appeal.