What is the difference between agents.txt and AGENTS.md?

agents.txt is a web-facing file at your domain root for autonomous AI agents to discover site capabilities. AGENTS.md is a repository-scoped markdown file primarily used in open-source projects to describe agent behavior to developers and coding assistants. agents.txt has an IETF standardization path; AGENTS.md is a community convention.

What is an A2A Agent Card (agent.json)?

An A2A Agent Card is a JSON file at /.well-known/agent.json that describes an agent's identity, capabilities, skills, and communication endpoints. It is part of Google's Agent-to-Agent (A2A) protocol, now under the Linux Foundation with 150+ member organizations.

Which agent discovery protocol should I use in 2026?

In April 2026, there is no single winner. For web-facing agent policies, agents.txt is the simplest starting point. For agent-to-agent communication, A2A Agent Cards have the strongest backing (Google + 150 orgs). For tool integration, MCP dominates with 97M monthly SDK downloads. A cross-protocol approach using multiple formats gives the broadest reach.

IETF agents.txt draft expires April 10, 2026

agents.txt vs AGENTS.md vs agent.json vs robots.txt

Q: What is the difference between agents.txt and robots.txt?

robots.txt tells crawlers what they cannot access (restrictions). agents.txt tells AI agents what they CAN do — endpoints, authentication, payment methods, and capabilities. robots.txt is a 30-year-old standard; agents.txt is an IETF draft (draft-srijal-agents-policy-00) expiring April 10, 2026.

8 competing approaches to agent discovery. No interoperability between them. Here is what each does, who backs it, and where it stands in April 2026.

robots.txt agents.txt AGENTS.md A2A Agent Cards (agent.json)MCP Server Cards Agent Web Protocol (AWP)AID (DNS TXT Records)DNS-AID (SVCB Records)

Side-by-Side Comparison

How the 8 agent discovery protocols compare across key dimensions.

Protocol	Format	Location	Status	Backer	Adoption
robots.txt	Plain text, line-based directives	`/robots.txt`	Standard	Google / W3C / Industry standard since 1994	Universal
agents.txt	Plain text with structured fields; JSON at /.well-known/agents	`/agents.txt and /.well-known/agents`	IETF Draft	IETF (draft-srijal-agents-policy-00)	Early
AGENTS.md	Markdown (.md)	`/AGENTS.md in repository root`	Community Draft	Pydantic AI community / Open-source convention	Early
A2A Agent Cards (agent.json)	JSON (Agent Card schema)	`/.well-known/agent.json`	Specification	Google + Linux Foundation (150+ organizations)	Growing
MCP Server Cards	JSON (SEP-1649 specification)	`Registry-hosted or self-declared`	Proposed	MCP community (Anthropic ecosystem)	Pre-adoption
Agent Web Protocol (AWP)	JSON at .well-known path	`/.well-known/agent.json`	Community Draft	Community / Independent developers	Early
AID (DNS TXT Records)	DNS TXT records	`_agent.example.com TXT record`	Proposed	Command Zero / ACDP working group	Pre-adoption
DNS-AID (SVCB Records)	DNS SVCB/HTTPS records	`_agent.example.com SVCB record`	Proposed	Community / DNS standards ecosystem	Pre-adoption

Which Protocol Should You Use?

You want to control crawler access

Use robots.txt. It is the universal standard for crawl permissions. Every search engine and AI crawler respects it. Start here.

You want agents to interact with your site

Use agents.txt. It declares endpoints, authentication, and capabilities. Simplest path for web-facing agent policies.

You are building agent-to-agent communication

Use A2A Agent Cards. Google + 150 organizations back this. Richest schema for agent identity and skills.

You are building tools for AI models

Use MCP and watch MCP Server Cards. 97M monthly SDK downloads make MCP the dominant tool integration layer.

You maintain an open-source AI project

Add an AGENTS.md to your repo root. It helps coding assistants and developers understand your agent's behavior.

You want maximum discoverability

Use all of them. A cross-protocol approach with robots.txt + agents.txt + A2A Agent Card covers the broadest surface. Global Chat indexes all formats.

Protocol Deep Dives

Detailed breakdown of each protocol: purpose, format, strengths, weaknesses, and code examples.

robots.txt

StandardUniversal

Controls which crawlers can access which parts of a website. Passive permission layer — tells bots what NOT to do.

Format

Plain text, line-based directives

Location

/robots.txt

Backed by

Google / W3C / Industry standard since 1994

Adoption

Universal — every major crawler respects it

Strengths

Universal adoption across all search engines and crawlers
Simple format that anyone can write
30 years of ecosystem support and tooling

Weaknesses

Only handles crawl permissions, not agent capabilities
No structured data about what agents CAN do
Cannot express authentication, payment, or API endpoints
Advisory only — no enforcement mechanism

Global Chat Support

Global Chat serves a robots.txt and reads robots.txt from discovered sites for crawl compliance.

Example

User-agent: *
Disallow: /private/
Allow: /public/

User-agent: GPTBot
Disallow: /

agents.txt

IETF DraftEarly

IETF draft expires April 10, 2026

Declares what AI agents CAN do on a site — endpoints, authentication, payment methods, and capabilities. The affirmative counterpart to robots.txt.

Format

Plain text with structured fields; JSON at /.well-known/agents

Location

/agents.txt and /.well-known/agents

Backed by

IETF (draft-srijal-agents-policy-00)

Adoption

Early — growing adoption among agent-first platforms

Strengths

Natural extension of the robots.txt mental model
IETF standardization path gives it institutional weight
Dual format: human-readable text + machine-readable JSON
Covers authentication, payment, and capability declaration

Weaknesses

IETF draft expires April 10, 2026 — renewal uncertain
No built-in agent-to-agent communication protocol
Adoption still limited to early movers

Global Chat Support

Full support. Global Chat validates agents.txt files, indexes them in the directory, and serves its own agents.txt.

Example

# agents.txt
User-agent: *
Agent-name: ExampleBot
Agent-description: Commerce agent
Auth-endpoint: /api/auth
Payment-methods: crypto, stripe
Capabilities: search, purchase, bid

AGENTS.md

Community DraftEarly

Human-readable markdown file describing an agent or repository to other AI coding agents. Primarily aimed at developer tooling and LLM-based code assistants.

Format

Markdown (.md)

Location

/AGENTS.md in repository root

Backed by

Pydantic AI community / Open-source convention

Adoption

Early — adopted in AI/ML open-source projects

Strengths

Human-readable and easy to write
No tooling required — just a markdown file
Fits naturally into repository-based workflows
Good for documenting agent behavior for developers

Weaknesses

No formal specification or schema
Not machine-parseable in a standardized way
Repository-scoped, not web-scoped — no URL-based discovery
No support for authentication, payment, or API endpoints

Global Chat Support

Global Chat can index AGENTS.md files from GitHub repositories and display them in the directory.

Example

# AGENTS.md

## Overview
This repository contains an AI agent for
code review automation.

## Capabilities
- Pull request analysis
- Code style enforcement
- Security vulnerability detection

## Integration
Supports GitHub Actions and GitLab CI.

A2A Agent Cards (agent.json)

SpecificationGrowing

Structured JSON card describing an agent's identity, capabilities, skills, and communication endpoints for Google's Agent-to-Agent protocol.

Format

JSON (Agent Card schema)

Location

/.well-known/agent.json

Backed by

Google + Linux Foundation (150+ organizations)

Adoption

150+ organizations — second-largest ecosystem after MCP

Strengths

Backed by Google and 150+ organizations under Linux Foundation
Rich schema covering identity, skills, and endpoints
Built-in support for agent-to-agent communication
Well-known URL convention enables web-scale discovery

Weaknesses

Complex schema — higher barrier to entry than text files
Tightly coupled to A2A protocol for full value
Not interoperable with MCP or agents.txt ecosystems

Global Chat Support

Global Chat indexes A2A Agent Cards and displays agent capabilities in the cross-protocol directory.

Example

{
  "name": "RecommendationAgent",
  "description": "Product recommendation agent",
  "url": "https://example.com/agent",
  "capabilities": {
    "streaming": true,
    "pushNotifications": false
  },
  "skills": [
    {
      "id": "product-search",
      "name": "Product Search"
    }
  ]
}

MCP Server Cards

ProposedPre-adoption

Capability metadata cards for MCP servers, enabling clients to discover what tools, resources, and prompts a server provides before connecting.

Format

JSON (SEP-1649 specification)

Location

Registry-hosted or self-declared

Backed by

MCP community (Anthropic ecosystem)

Adoption

Pre-adoption — spec work active, 97M monthly MCP SDK downloads provide potential reach

Strengths

Builds on MCP, the dominant tool integration protocol (97M downloads/month)
Would enable pre-connection capability discovery
Active spec work in SEP-1649 with community engagement

Weaknesses

Still in proposal stage — not yet part of MCP spec
MCP is tool-integration focused, not agent-discovery focused
No standardized hosting location yet
Does not cover agent-to-agent scenarios

Global Chat Support

Global Chat tracks MCP Server Cards development and will index them once the specification is finalized.

Example

{
  "name": "github-mcp-server",
  "version": "1.2.0",
  "description": "GitHub integration",
  "tools": [
    "create_issue",
    "search_repos",
    "create_pull_request"
  ],
  "resources": ["repo://owner/name"],
  "auth": "oauth2"
}

Agent Web Protocol (AWP)

Community DraftEarly

Web-native agent discovery using standard .well-known URLs. Defines how agents advertise capabilities through HTTP endpoints.

Format

JSON at .well-known path

Location

/.well-known/agent.json

Backed by

Community / Independent developers

Adoption

Early — limited implementations

Strengths

Uses existing web infrastructure (.well-known)
HTTP-native — no new protocols needed
Compatible with existing web security models

Weaknesses

Overlaps with A2A Agent Cards on .well-known/agent.json
No major corporate backer
Limited tooling and validator support

Global Chat Support

Global Chat can discover AWP endpoints and index them alongside other protocols.

Example

// GET /.well-known/agent.json
{
  "agent": "TaskRunner",
  "version": "1.0",
  "endpoints": [
    "/api/tasks",
    "/api/status"
  ],
  "auth": "bearer"
}

AID (DNS TXT Records)

ProposedPre-adoption

DNS-based agent identity and discovery using TXT records. Allows domain owners to declare agent capabilities at the DNS layer.

Format

DNS TXT records

Location

_agent.example.com TXT record

Backed by

Command Zero / ACDP working group

Adoption

Pre-adoption — concept stage

Strengths

Leverages existing DNS infrastructure — no new servers needed
Domain-level authority — only domain owners can set records
Familiar pattern (like SPF/DKIM for email)

Weaknesses

DNS TXT records have size limits (255 bytes per string)
Cannot express rich capability schemas
DNS propagation delays make updates slow
Requires DNS access, which many developers lack

Global Chat Support

Global Chat monitors AID proposals and will support DNS-based discovery when standards mature.

Example

; DNS TXT record
_agent.example.com. IN TXT
  "v=agent1; name=SupportBot;
   cap=search,chat;
   url=https://example.com/agent"

DNS-AID (SVCB Records)

ProposedPre-adoption

Advanced DNS-based agent discovery using SVCB/HTTPS resource records. Richer than TXT records, supports structured parameters.

Format

DNS SVCB/HTTPS records

Location

_agent.example.com SVCB record

Backed by

Community / DNS standards ecosystem

Adoption

Pre-adoption — requires SVCB support in resolvers

Strengths

SVCB records support structured key-value parameters
Built on IETF-standard record type (RFC 9460)
Can encode priority, transport, and endpoints natively

Weaknesses

SVCB support still rolling out across DNS resolvers
Complex to configure compared to file-based approaches
No existing tooling for agent-specific SVCB records
Requires DNS expertise that most developers lack

Global Chat Support

Global Chat tracks DNS-AID development as part of the discovery landscape monitoring.

Example

; DNS SVCB record
_agent.example.com. IN SVCB 1
  agent.example.com.
  alpn=h2 port=443
  key65400=search,chat

The Fragmentation Reality

In April 2026, 8+ protocols compete to define how AI agents discover each other. None of them are interoperable. An agent that reads agents.txt cannot parse A2A Agent Cards. An MCP client has no way to discover agents using DNS TXT records.

The IETF agents.txt draft expires on April 10, 2026. Whether it gets renewed will signal the standards body's commitment to file-based agent policy. Meanwhile, Google's A2A with 150+ organizations has the largest coalition, and MCP with 97M monthly downloads has the largest developer base.

The path forward is not picking one winner. It is building infrastructure that spans all of them — a discovery layer that reads every format and translates between protocols. That is what Global Chat is building.

Make your agent discoverable across all protocols

Global Chat indexes agents.txt, A2A Agent Cards, MCP servers, and more. Validate your discovery files or register your agent in the cross-protocol directory.

Validate agents.txt Browse Directory Register Agent