Layers

Agent standards and conventions are easier to compare when grouped by the role they play in an agent workflow. Use this map to decide which artifact belongs in a repository, docs site, or integration surface.

Quick Decision Map

Need	Start with	Why
Tell coding agents how to work in a repository	AGENTS.md	Cross-tool project instructions are the most useful baseline.
Preserve long-lived project context	MEMORY.md or Memory Bank	These files keep stable knowledge and task state outside a single chat.
Package a reusable agent capability	SKILL.md	Skills are loaded on demand when a task matches the description.
Store prompts as versioned assets	.prompty, .prompt, or system_prompt.txt	Prompt files keep model instructions inspectable and reviewable.
Restrict files or stage risky work	.aiignore, scoped rules, and PLAN.md	Guardrails reduce accidental reads, writes, and unreviewed execution.
Make behavior measurable	EVAL.yaml	Evals turn agent quality into a repeatable check.
Trace agent execution	OpenTelemetry GenAI or OpenInference	Semantic conventions make model, agent, and tool telemetry comparable.
Help LLMs discover public docs	llms.txt and llms-full.txt	Predictable Markdown entry points reduce scraping ambiguity.
Connect an agent to tools and data	MCP	MCP standardizes the agent-to-tool boundary.
Connect independent agents	A2A	A2A standardizes discovery and task exchange between agents.
Connect a coding agent to an editor	ACP	ACP standardizes the client-to-agent boundary.

Instruction Layer

These files tell agents how to behave in a repository.

• AGENTS.md
• CLAUDE.md
• GEMINI.md
• Tool-specific rules such as Cursor rules, Cline rules, and Copilot instructions
• .aiignore

Context and State Layer

These files preserve project knowledge, task state, or execution intent.

• MEMORY.md
• Memory Bank folders such as cline_docs/ and .roo/
• PLAN.md

Prompt and Capability Layer

These files package model instructions and agent capabilities into reusable assets.

• .prompty
• .prompt
• system_prompt.txt
• SKILL.md

Execution and Safety Layer

Runtime loops, authorization checks, budgets, retries, and output validation are usually implemented by the agent host. Repository conventions can influence these controls but do not implement the runtime by themselves.

• .aiignore
• PLAN.md
• Tool-specific scoped rules

Evaluation Layer

These files make behavior testable and repeatable.

• EVAL.yaml

Observability Layer

These semantic conventions describe agent runs as traces, spans, metrics, and events.

• OpenTelemetry GenAI Semantic Conventions
• OpenInference Semantic Conventions

Discoverability Layer

These files expose structured information to LLMs and API-aware clients.

• llms.txt
• llms-full.txt
• pricing.md
• auth.md
• /.well-known/ai-plugin.json

Interoperability Layer

These standards define interoperability across different harness boundaries.

• Model Context Protocol for agents, tools, and data
• Agent2Agent Protocol for independent agents
• Agent Client Protocol for coding agents and editor clients