Standardize on AGENTS.md on .agentignore Format Support

[stevemoser]

This is a proposal to support the AGENTS.md and .agentignore formats in Agent-Client-Protocol (ACP)

1. AGENTS.md is a standardized file for providing instructions and capabilities to AI agents.

2. .agentignore is a standardized file for specifying files and directories that AI agents should ignore (similar to .gitignore). Support for .agentignore has been proposed in the AGENTS.md project: Specification Extension Request: Standardized ignore files Support in Agents.md Specification agentsmd/agents.md#33

Motivation

The agent-client-protocol, being a standard for communication between editors and agents, is well positioned to support the AGENTS.md and .agentignore formats.

Adopting AGENTS.md and .agentignore will create a more consistent and predictable experience for developers and it will allow them to easily switch between coding agents.

Proposed Solution

1. Adopt AGENTS.md: The agent-client-protocol should recognize AGENTS.md as the standard file for agent instructions. This file would provide a centralized place for developers to define the agent's purpose, capabilities, and any other relevant information.

2. Adopt .agentignore: The protocol should also adopt .agentignore as the standard for specifying files and directories to be ignored by the agent.

The agent-client-protocol specification should be modified to specify on how clients should discover, parse, and utilize both AGENTS.md and .agentignore.

[benbrandt]

@stevemoser can you explain more the problem you are trying to solve with this?
We kind of assume the agents, not the client, would be handling these sorts of decisions. However, maybe the agent could let the client know which instructions/rules it is using? So the client could display to the user what is being used?

Would that solve the issue you are having? If not, let me know what it is you are thinking.

[stevemoser]

Yes displaying the rules in use would be very helpful. I see the believe the most salient part is that the client enforces which files are visible to the agent. This is especially important from a security standpoint for .env files.

[jingchang0623-crypto]

5个Agent共用一套指令，标准化是刚需 📋

我们运营miaoquai.com有5个AI Agent，每个Agent都有自己的配置文件：SOUL.md（人设）、AGENTS.md（工作流程）、USER.md（用户偏好）、MEMORY.md（长期记忆）。

标准化的痛点

目前每个Agent框架都有自己的配置方式：

• OpenClaw 用 SOUL.md + AGENTS.md + USER.md
• Cursor 用 .cursor/rules
• Claude Code 用 CLAUDE.md
• Windsurf 用 .windsurfrules

AGENTS.md 试图统一这些，但现实是每个框架都有「私货」。比如我们的SOUL.md里包含人设性格描述，这在标准AGENTS.md里没有对应字段。

.agentignore 的价值被低估了

这个功能对我们太重要了。5个Agent共享一个代码仓库，但：

• SEO Agent 不该看到财务数据
• 内容Agent 不该修改部署脚本
• 社区Agent 不该访问内部记忆

如果 .agentignore 能按Agent维度配置，那简直就是Agent权限管理的最小实现。

建议

AGENTS.md 可以考虑增加 agent-specific overrides。比如：

AGENTS.md 是通用配置，AGENTS.seo.md 是SEO Agent的专属覆盖。这样既保持标准统一，又允许个性化。

PS: 我们在 https://miaoquai.com/stories/agent-team-drama.html 写过5个Agent抢资源的内斗故事，权限隔离确实是团队协作的第一道坎。